Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

692 https://internationalpubls.com 

MDVPB: Design of an Efficient Modified Densenet-161 Model for 

Advanced Vein Pattern Recognition in Biomedical Systems 

 

Prof Manisha A.Gawande1, Dr Suchita W Varade2 
1Assistant Professor, Electronics and Telecommunication Engineering, Sipna College of Engineering and Technology, 

Amravati, Maharashtra, India 

2Professor, Electronics Engineering, Priyadarshini College of Engineering, Nagpur, Maharashtra, India 

1Email: manishagawande8587@gmail.com 

 

Article History: 

Received: 30-05-2024 

Revised: 12-07-2024 

Accepted: 29-06-2024 

Abstract:  

There is an ever-increasing need for an efficient, and accurate vein detection system that can 

work in clinical scenarios. Traditional methods, while effective, grapple with challenges like 

high error rates and susceptibility to environmental variations, making the need for advanced 

solutions more pressing. This study addresses these limitations by unveiling a 

groundbreaking Convolutional Neural Network (CNN) architecture, tailored specifically for 

the nuanced recognition of vein patterns, a critical component in biomedical systems. The 

existing frameworks in vein pattern recognition often falter in handling intricate variations, 

leading to diminished accuracy and reliability levels. The proposed model revolutionizes this 

landscape by integrating a modified Densenet-161 structure for analyzing blood flow velocity 

and recognizing dorsal hand vein locations. The ingenuity lies in the removal of the 

conventional classification layer, transforming it into a robust feature embedder. This 

strategic alteration enables the extraction of remarkably distinct vein features from hand 

images, a leap forward in capturing the unique identifiers crucial in biomedical systems. The 

methodology employed is twofold: a meticulous training and validation phase, followed by 

a testing phase under different verification scenarios. The training harnesses the power of the 

CNN feature embedder, extracting and classifying features with an unprecedented precision. 

The testing phase introduces a dual input system, comparing known and unknown vein 

patterns using Euclidean distance, a method that quantifies similarity with remarkable 

accuracy levels. This process significantly reduces intra-class variations while amplifying 

inter-class differences, culminating in a system that is not only highly accurate but also 

reliable for clinical scenarios. The model's superiority is further evidenced by its performance 

on various datasets, showcasing improvements in precision, accuracy, recall, AUC, and 

reduced delay, outperforming existing methods by notable margins. The model can thus be 

used for intravenous catheterization, blood check-up, and analysis of different veins 

identified related to different diseases. 

Keywords: Biomedical Identification, Convolutional Neural Network, Vein Pattern 

Recognition, Feature Extraction, Biomedical Security, Scenarios. 

 

1. Introduction 

The evolution of biomedical identification systems has been a focal point in the realm of security and 

personal identification. Among various biomedical traits, vein patterns have emerged as a distinctive 

and secure modality, owing to their uniqueness and resistance to external influences. However, the 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

693 https://internationalpubls.com 

journey to perfection in vein pattern recognition technology is fraught with complexities and 

challenges, necessitating continual advancements and innovative approaches. 

Historically, vein pattern recognition systems have employed a range of techniques, from simplistic 

pattern matching to sophisticated imaging technologies. Despite these advancements, the field has 

consistently grappled with challenges such as high error rates, susceptibility to environmental 

variations, and the need for robustness against spoofing attempts. These challenges underscore the 

pressing need for a more sophisticated and reliable method of vein pattern recognition. 

Enter the world of Convolutional Neural Networks (CNNs), a breakthrough in machine learning that 

has revolutionized image processing and pattern recognition. CNNs, with their deep learning 

capabilities, have shown remarkable success in various domains, including biomedical identification. 

However, the application of CNNs in vein pattern recognition has not been fully explored, particularly 

in terms of optimizing architecture for enhanced feature extraction and accuracy levels. 

The proposed research aims to fill this gap by introducing a novel CNN architecture, meticulously 

designed for vein pattern recognition in biomedical systems. This innovative model leverages the 

strengths of Densenet-161, a proven CNN structure, but with a critical modification – the removal of 

its classification layer. This transformation allows the model to function as a sophisticated feature 

embedder, adept at extracting nuanced features from hand vein images & samples. This approach is 

not just a technical enhancement but a strategic shift in how vein patterns are analyzed and recognized 

for different scenarios. 

The model’s functionality is showcased in two distinct phases: training (and validation) and testing 

(verification scenario). The training phase utilizes the CNN feature embedder to process vein input 

images, extracting and classifying features with an advanced level of precision. The testing phase, on 

the other hand, adopts a dual-input scheme. Here, the model processes both a known (enrollment input) 

and an unknown (query input) vein pattern, comparing the extracted features to ascertain their 

similarity. This comparison, based on Euclidean distance, culminates in a score that effectively 

quantifies the resemblance between the two feature sets, thereby determining the authenticity of the 

biomedical input. 

The architecture's superiority is further validated through extensive testing across multiple datasets, 

where it demonstrated remarkable improvements in various key metrics. Notably, it exhibited 

significant enhancements in precision, accuracy, recall, and AUC, while also reducing delay compared 

to existing methods. This research thus presents not just an optimized CNN model but a comprehensive 

solution to the challenges that have long plagued vein pattern recognition in biomedical systems. 

In essence, this study not only proposes a groundbreaking model but also paves the way for a new era 

in biomedical identification, one that promises greater security, reliability, and accuracy. The 

introduction section of this paper seeks to lay down the foundation of this research, detailing the 

background, challenges, and the novel approach undertaken, setting the stage for a detailed exploration 

of the model and its implications in the succeeding sections. 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

694 https://internationalpubls.com 

Motivation & Contribution 

In the dynamic landscape of biomedical identification, the quest for a robust and foolproof system 

remains a paramount objective. Vein pattern recognition, a relatively newer entrant in the biomedical 

sphere, offers a promising avenue owing to its inherent security features and uniqueness. However, 

the journey to harnessing its full potential is not devoid of hurdles. The motivation for this research 

emerges from this very pursuit of overcoming the existing limitations in vein pattern recognition and 

pushing the boundaries of what is possible with current technologies. 

The primary motivation driving this study is the inherent challenges that plague current vein pattern 

recognition systems. These include high error rates, sensitivity to environmental factors like lighting 

and temperature, and vulnerability to spoofing. These challenges not only compromise the security 

and reliability of biomedical systems but also limit their widespread adoption. Additionally, the lack 

of a specialized architecture in existing CNN models for effectively capturing the intricacies of vein 

patterns further motivates the need for innovation in this domain. 

This research, therefore, aims to address these challenges head-on by introducing a novel CNN-based 

model designed specifically for vein pattern recognition. The model's architecture, a modified version 

of the Densenet-161 structure, is a strategic response to the need for a more efficient feature extraction 

mechanism in vein pattern recognition. By removing the classification layer and repurposing the model 

as a feature embedder, this research makes a leap forward in capturing the complex and unique features 

of vein patterns. 

The contributions of this study are manifold and significant. Firstly, it introduces a novel CNN 

architecture that is specifically optimized for vein pattern recognition, a feat that has not been 

extensively explored in previous research. This model demonstrates an enhanced ability to minimize 

intra-class variations while maximizing inter-class differences, resulting in a more accurate and secure 

biomedical identification system. 

Secondly, the research methodology employed, including the dual-input scheme in the verification 

scenario, sets a new standard in the field. This method not only enhances the accuracy of the 

recognition process but also adds an extra layer of security by effectively distinguishing between 

genuine and impostor vein patterns. 

Moreover, the empirical results obtained from multiple dataset evaluations provide concrete evidence 

of the model's superiority over existing methods. The significant improvements in precision, accuracy, 

recall, AUC, and reduced delay are not just numbers but represent a substantial leap in the effectiveness 

of biomedical identification systems. 

In essence, this research makes a pivotal contribution to the field of biomedical identification. It not 

only presents a groundbreaking CNN architecture for vein pattern recognition but also lays the 

groundwork for future advancements in this domain. The Motivation and Contribution section of this 

paper delves into these aspects, underscoring the rationale behind this research and the significant 

strides it makes in enhancing the security and reliability of biomedical systems. 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

695 https://internationalpubls.com 

2. Literature Review 

The literature review on finger vein recognition technologies demonstrates significant progress in the 

field, leveraging advanced methodologies for enhanced authentication and identification systems. This 

section synthesizes the contributions from recent studies, highlighting their objectives, methodologies, 

and findings within the broader context of biological authentications. 

Krishnan and Thomas [1] introduced a novel approach for finger vein recognition by focusing on the 

anatomical features of vein patterns. Their method, termed FEBA classification, harnesses the unique 

aspects of vein anatomy for pattern recognition and matching, significantly improving the accuracy of 

finger vein biological s. This study lays the groundwork for utilizing anatomical uniqueness in 

biological authentication, presenting a detailed analysis of feature extraction and representation 

techniques that enhance the matching process. 

Building upon the advancements in feature extraction, Song et al. [2] developed EIFNet, a network 

that fuses explicit and implicit features for finger vein verification. Their research underscores the 

importance of deep learning in extracting and combining various features of vein patterns, thereby 

achieving a higher verification accuracy. EIFNet exemplifies the integration of convolutional networks 

in biological s, illustrating the potential of feature fusion in improving the reliability of finger vein 

verification systems. 

In a different vein, Xu et al. [3] explored the medical application of vein imaging, specifically for 

portal vein pressure estimation and hypertension discrimination. Their study utilizes subharmonic 

scattering of ultrasound contrast agent microbubbles, offering a non-invasive method for blood 

pressure estimation. Although focused on medical diagnostics, this research contributes to the broader 

understanding of vein imaging and its potential applications, including biological s. 

He et al. [4] addressed the challenge of image quality in finger vein recognition with the development 

of Neighbors-Based Binary-GAN (NB-GAN) for image deblurring [4]. This technique mitigates 

texture loss and enhances image clarity, facilitating more accurate vein pattern recognition. The 

application of GANs in improving image quality underscores the intersection of image processing and 

biological authentication, highlighting the importance of high-quality inputs for reliable identification 

systems. 

Yang et al. [5] presented a study on small-area finger vein recognition, focusing on the challenges and 

solutions for implementing finger vein biological s with embedded sensors [5]. Their approach, based 

on dictionary learning, adapts to the constraints of small-area imaging, demonstrating the feasibility 

of compact and efficient finger vein recognition systems suitable for mobile and embedded 

applications. 

Further research by Zhao et al. [6] into single-sample recognition via competitive and progressive 

sparse representation emphasizes the advances in handling limited data scenarios [6]. Their 

methodology showcases the effectiveness of sparse representation techniques in enhancing the 

accuracy of finger vein recognition from a single sample, addressing a common limitation in biological 

systems. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

696 https://internationalpubls.com 

Cross-dataset recognition, a critical challenge in the domain, is tackled by Huang and Guo [7], who 

propose methods for domain adaptation and alignment in finger vein recognition with single-source 

data [7]. Their work on fast transfer learning and domain adaptation strategies marks a significant step 

toward overcoming dataset variability, enhancing the robustness and scalability of finger vein 

recognition systems. 

Li and Zhang [8] introduced FV-ViT, a vision transformer model tailored for finger vein recognition 

[8]. This study illustrates the applicability of transformers in biological authentication, leveraging their 

capability for feature extraction and computational efficiency. The adoption of vision transformers in 

finger vein recognition highlights the ongoing evolution of neural network architectures in the field. 

Tong et al. [9] developed a lightweight network focusing on contextual and morphological awareness 

for hepatic vein segmentation [9]. Although primarily aimed at medical imaging, the principles of 

attention mechanisms and lightweight design are relevant to biological authentication, offering insights 

into efficient and accurate vein imaging technologies. 

Lastly, Huang, Ma, and Wang [10] explored axially enhanced local attention networks for finger vein 

recognition, further advancing the application of attentional mechanisms in feature extraction [10]. 

Their work demonstrates the benefits of deep learning in capturing intricate details of vein patterns, 

contributing to the ongoing enhancement of recognition accuracy and system performance. 

Li et al. [16] introduced a novel approach for finger vein feature extraction by combining joint 

discriminative analysis with low-rank projection. Their method emphasizes the importance of jointly 

embedding features in a low-dimensional space while retaining discriminative properties for accurate 

recognition. This technique addresses the challenge of domain adaptation in biological systems, 

showcasing the potential of low-rank representation in enhancing the robustness of finger vein 

recognition. 

Zhang et al. [17] leveraged the power of ResNet with self-attention mechanisms for finger vein 

recognition. Their study highlights the efficacy of deep learning models, specifically convolutional 

neural networks (CNNs), augmented with self-attention to improve feature extraction and learning 

rates dynamically. This research demonstrates the ongoing integration of advanced neural network 

architectures and attention mechanisms in biological s, offering significant improvements in 

recognition accuracy. 

Garcia-Martin and Sanchez-Reillo [18] explored the application of vision transformers in vein 

biological recognition, comparing their performance with conventional CNNs. Their work provides 

insight into the advantages of transformers over CNNs in certain scenarios, emphasizing the role of 

transfer learning in enhancing model effectiveness. This study broadens the understanding of machine 

learning algorithms' applicability in biological recognition, particularly in mobile and contactless 

contexts. 

Pan et al. [19] developed the Disentangled Representation and Enhancement Network (DRE-Net) for 

vein recognition. Their approach employs a multi-scale attention residual block and a weight-guided 

feature enhancement module to improve recognition performance. This work illustrates the potential 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

697 https://internationalpubls.com 

of disentangled representations and attention mechanisms in accurately capturing vein patterns, 

offering advancements in feature extraction methodologies. 

Güldenring et al. [20] focused on the fine-grained analysis of plant structures, including leaves, stems, 

and veins, through robotics and automation. Although not directly related to human biological s, this 

research contributes to the broader field of pattern recognition and feature extraction, offering insights 

into complex biological structure analysis. 

In the realm of palm vein authentication, Ma et al. [21] introduced focal contrastive learning to enhance 

feature extraction and recognition accuracy. Their method focuses on hard sample mining, 

demonstrating the effectiveness of contrastive learning in biological technologies. This study 

underlines the importance of innovative learning strategies in overcoming challenges related to intra-

class variability and sample difficulty. 

Li et al. [22] tackled the security aspect of biological systems by developing a transformer-based 

defense GAN against palm-vein adversarial attacks. Their work addresses the vulnerability of 

biological authentication systems to adversarial perturbations, proposing a robust defense mechanism. 

This research underscores the critical need for security in biological systems, emphasizing the role of 

adversarial learning and defense strategies. 

Qin et al. [23] proposed a multiscale transformer for palm-vein recognition, incorporating label 

enhancement techniques. Their study showcases the use of transformers in handling the complexity of 

vein patterns, enhancing recognition accuracy through multiscale feature extraction. This contribution 

highlights the adaptability of transformer models in biological recognition, paving the way for more 

accurate and efficient authentication systems. 

Zhang et al. [24] presented CECNet, a neural network for palm vein recognition that utilizes coordinate 

encoding and a novel loss function. Their approach demonstrates the potential of competitive neural 

networks and adaptive feature extraction methods in improving recognition performance. This 

research contributes to the ongoing evolution of neural network architectures and loss functions 

tailored for biological recognition. 

Nguyen et al. [25] introduced LAWNet, a lightweight attention-based model for wrist vein verification 

on smartphones using RGB images. Their work emphasizes the importance of attention mechanisms 

in enhancing feature extraction from RGB images, showcasing the feasibility of deploying advanced 

biological systems on mobile devices. This study illustrates the potential for widespread application 

of vein recognition technologies, facilitated by advances in deep learning and mobile computing. 

Collectively, these studies illustrate the dynamic and evolving nature of vein recognition research, 

incorporating sophisticated machine learning algorithms, attention mechanisms, and novel feature 

extraction techniques to tackle the challenges of biological authentication. The integration of these 

advanced methodologies not only enhances the accuracy and security of vein recognition systems but 

also broadens their applicability across various platforms and devices for different scenarios. As the 

field continues to advance, the emphasis on innovation, security, and user convenience will likely drive 

further developments in biological authentication technologies. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

698 https://internationalpubls.com 

3. Design of the Proposed Model 

To overcome issues of low efficiency & high complexity which are present in the currently reviewed 

vein analysis methods, this section discusses design of an intricately designed architecture of the 

Modified Densenet-161 Model for Advanced Vein Pattern Recognition in Biomedical Systems. The 

convolutional neural network (CNN) serves as the foundation, ingeniously tailored to sift through the 

complex visual fusion of vein patterns with unmatched precision levels. As per figure 1.1, at the heart 

of this architectural marvel lies the DenseNet 161 framework, renowned for its dense connectivity 

pattern that fosters information flow between layers, thereby enhancing feature extraction without the 

heft of computational redundancy. This model transcends traditional paradigms by incorporating a 

dual-input process, a novel approach that fuses pairs of vein images, harnessing the power of 

comparative analysis to distill and recognize the intricate nuances of vein patterns. Beyond the core of 

CNN and DenseNet 161, the architecture unfolds into a constellation of meticulously crafted blocks, 

each serving a pivotal role—from preprocessing that standardizes and primes the images for analysis, 

to feature embedding which eschews the conventional classification layer for a more potent 

representation of vein features. This ensemble of blocks operates in parallel, executing a symphony of 

complex data processing tasks with the agility and precision that the realm of vein pattern recognition 

demands. Through this integration of advanced technologies and innovative methodologies, the 

MDVPR model emerges as a beacon of progress, setting new benchmarks in the accuracy, efficiency, 

and reliability of biological identification systems. 

In the ambit of advancing biometric recognition technologies, the development of the Modified 

Densenet-161 Model for Advanced Vein Pattern Recognition (MDVPR) represents a paradigmatic 

shift for the process. This model intricately weaves together convolutional neural networks (CNNs) 

with a reimagined DenseNet-161 architecture, specifically optimized for the intricate task of analyzing 

blood flow velocity and identifying dorsal hand vein locations. Central to this innovation is the 

strategic removal of the conventional classification layer, which is traditionally pivotal in CNN 

architectures for the final classification of features. Instead, this layer is metamorphosed into a robust 

feature embedder, a transformation that is both novel and critical for different image sets. This 

alteration facilitates the extraction of vein features with exceptional distinctness, thereby enhancing 

the model's ability to discern the unique identifiers embedded within hand images, crucial for 

applications within biomedical systems. 

The foundational operations that encapsulate the essence of this model commence with the input image 

sets, represented as X, which undergo a preprocessing phase to ensure uniformity in scale and contrast, 

facilitating optimal extraction of features. The feature extraction process within the DenseNet-161 

architecture is mathematically represented via equation 1, 

𝐹(𝑥) = 𝐻𝑙([𝑥0, 𝑥1, . . . , 𝑥𝑙 − 1]) … (1) 

Where, Hl represents the composite function of operations (Batch Normalization, ReLU, Convolution) 

applied at the lth layer which is represented via equation 2, and [x0,x1,...,xl−1] represents the 

concatenation of the feature maps produced by all preceding layers.  

𝐻𝑖(𝑥) = 𝜎 (𝐵𝑁(𝐶𝑜𝑛𝑣(𝑥))) … (2) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

699 https://internationalpubls.com 

The modification of the DenseNet-161, to serve as a feature embedder, introduces a mapping function 

M:F(x)→V, where V signifies the vein feature vectors that depend on blood flow & other parameter 

sets.  

 

Figure 1.1. Model Architecture for the Proposed Vien Detection Process 

This mapping is crucial, transforming the high-dimensional feature space F(x) into a more compact, 

yet informative, representation space V, optimized for vein pattern recognition, which are represented 

via equation 3, 

𝑉 = 𝑀(𝐹(𝑥)) = 𝑊 ⋅ 𝐹(𝑥) + 𝑏 … (3) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

700 https://internationalpubls.com 

Where, W and b represent the weights and bias of a fully connected layer designed to highlight vein 

patterns within the feature map F(x) sets. The essence of M(x) lies in its ability to prioritize spatial 

features that are indicative of vein locations, thus facilitating accurate localizations. The process of 

vein localization from the vein-specific representation V involves identifying regions within V that 

exceed a certain threshold θ, indicative of vein presence, and is represented via equation 4, 

𝐿(𝑉) = { 𝑝 ∈ 𝑉 ∣∣ 𝑝 > 𝜃 } … (4) 

Where, L(V) represents the set of localized vein points in V, and p represents individual points in V 

sets. The threshold θ is determined based on the distribution of values in V that correspond to vein 

patterns, optimized through empirical analysis. 

In this context, the process of recognizing vein regions from the embedded features involves a 

verification phase, which can be modeled as a comparison between the extracted feature vector V and 

a pre-defined threshold θ, to determine the presence of specific vein patterns. This comparison is 

formulated via equation 5, 

𝑅(𝑉) = {1 𝑖𝑓 𝑑(𝑉, 𝑉′) ≤ 𝜃, 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 … (5) 

Where, R(V) represents the recognition result, d(V, V′) is a distance metric (Euclidean distance) 

measuring the similarity between the feature vector V and a reference vector V′, and θ represents the 

decision threshold levels. During meticulous training and validation phase of the model utilizes a loss 

function designed to minimize the difference between the predicted vein patterns and the actual vein 

patterns, ensuring precise feature classification, which is represented via equation 6, 

𝐿 =
1

𝑁
∑(𝑦𝑖 − 𝑦′𝑖)2 … (6) 

Where, N represents the number of training samples, yi the true label, and y’i the predicted label for 

the ith sample sets. The gradient descent method, applied to optimize the model's parameters, employs 

derivatives of the loss function with respect to the model's weights, W, to iteratively adjust and improve 

the model’s performance via equation 7, 

𝛥𝑊 = −𝜂𝛻𝐿(𝑊) … (7) 

Where, η represents the learning rate, and ∇L(W) is the gradient of the loss function with respect to the 

weights. Based on this, the results of Vien Detection can be observed from figure 1.2 as follows, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

701 https://internationalpubls.com 

 

Figure 1.2. Sample Results of the Modified DenseNet Process (Input Image, Vein Detected Image, 

Fine Tuned Image Samples) 

Next, the model’s efficiency is tested using an intricate application of biometric recognition, 

particularly within the realm of vein pattern identification operations. This comparison leverages the 

mathematical properties of Euclidean distance, encapsulated via equation 8, 

𝐷(𝑣𝑖, 𝑣𝑗) = √∑(𝑣𝑖𝑘 − 𝑣𝑗𝑘)2 … (8) 

Where, D(vi,vj) represents the Euclidean distance between the feature vectors vi and vj of the i-th and 

j-th vein images, respectively, across an n-dimensional set of feature spaces. This methodological 

choice is pivotal, as it provides a quantifiable measure of similarity with astonishing precision, 

enabling the system to discern minute discrepancies between vein patterns with high efficiency levels. 

The dual input process, at its core, is a fusion of mathematical operations, each step is designed to 

reduce intraclass variations while simultaneously magnifying the distinctions between different 

classes. At the outset, each detected vein region undergoes a transformation into a feature vector via 

equation 9, 

𝐹(𝑋) = ∫ 𝜙(𝑥)𝑑𝑥
𝛺

0

… (9) 

Where, F(X) delineates the feature vector derived from the detected vein region X, ϕ(x) is the 

embedding function mapping each pixel to a high-dimensional feature space, and Ω represents the 

domain of image samples. This transformation is crucial, as it encapsulates the essence of the vein 

pattern into a compact, computationally tractable form for different scenarios. 

Upon the extraction of feature vectors, the dual input system commences its comparison, employing 

the Euclidean distance as a metric to evaluate the similarity between known and unknown vein 

patterns. This evaluative process is further refined through the introduction of a thresholding 

mechanism, defined via equation 10, 

𝑇 = 𝜇 + 𝛼 ⋅ 𝜎 … (10) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

702 https://internationalpubls.com 

Where, T signifies the threshold value, μ is the mean Euclidean distance observed across all 

comparisons, σ represents the standard deviation, and α is a scaling factor determined empirically by 

the evaluation process. This threshold serves as a decisive boundary, delineating matches from 

mismatches based on the calculated distances. For vein regions adjudged similar, the model applies a 

marking process, encapsulated by a function represented via equation 11,  

𝑀(𝑋, 𝑌) = 𝛿(𝐷(𝑋, 𝑌) < 𝑇) … (11) 

Where, M(X,Y) is the binary marking function indicating whether the vein regions X and Y are 

considered a match (1 for match, 0 for no match), and δ is the indicator process. This marking is 

instrumental in the identification of vein regions with higher efficiency levels, facilitating their 

prioritization in subsequent analyses or clinical applications. The culmination of this process is the 

output of marked vein regions, characterized by higher efficiency levels, through an integrative 

operation represented via equation 12, 

𝑂 = ∑ 𝑀(𝑉𝑖, 𝑉𝑟𝑒𝑓)

𝑁

𝑖=1

… (12) 

Where, O represents the output comprising all marked vein regions from the set of N detected regions 

Vi, and Vref represents the reference vein patterns. This process embodies the essence of the dual input 

process, synthesizing the comparative analysis into actionable insights, thereby marking a paradigm 

shift in the precision and reliability of vein pattern recognition systems. Through this sophisticated 

fusion of mathematical formulations and computational processes, the proposed model not only 

elevates the accuracy of vein pattern identification but also enhances its applicability in clinical 

scenarios, promising a future where biometric recognition systems achieve unprecedented levels of 

efficiency and reliability levels. Performance of this model was estimated in terms of different 

evaluation metrics, and compared with existing methods in the next section of this text. 

4. Result Analysis 

In advancing biological and biomedical identification systems, the proposed model stands as a paragon 

of innovation, incorporating a reimagined Densenet-161 architecture meticulously tailored for the 

nuanced task of blood flow velocity analysis and dorsal hand vein location recognition. This model 

eschews the traditional classification layer found in conventional architectures, opting instead for a 

transformation into a potent feature embedder—a move that exemplifies strategic foresight. This 

fundamental reconfiguration facilitates the extraction of vein features with unparalleled distinctness 

from hand images, marking a significant departure from previous methodologies. The process 

underpinning this breakthrough involves a rigorous bifurcated methodology: initially, the model 

undergoes a thorough phase of training and validation, designed to hone its capabilities by learning 

from a diverse array of hand images to accurately capture the intricate vein patterns. Subsequently, the 

model is subjected to a testing phase, adopting a verification scenario that meticulously evaluates its 

proficiency in identifying and distinguishing vein patterns. This sophisticated approach not only 

underscores the model's ability to discern highly specific biological data but also its adaptability in 

processing and analyzing complex datasets, thereby setting a new benchmark in the domain of vein 

pattern recognition technology process. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

703 https://internationalpubls.com 

The experimental setup for evaluating the Modified Densenet-161 Model for Advanced Vein Pattern 

Recognition in Biomedical Systems (MDVPR) is meticulously designed to assess its performance 

across various dimensions, including precision, accuracy, recall, delay, AUC, and specificity. This 

section details the experimental framework, datasets used, and the parameters set for the evaluation, 

aiming to ensure reproducibility and transparency in assessing MDVPR's efficacy in vein pattern 

recognition. 

Contactless Knuckle-Palm Print & Vein Dataset: This dataset comprises images of the knuckle-

palm print and associated vein patterns captured in a contactless manner. It includes 2,400 images from 

200 subjects, with each subject contributing 12 images. The images are captured under varied lighting 

conditions to simulate real-world scenarios. Each image has a resolution of 1920x1080 pixels, 

providing high-definition quality for detailed vein pattern analysis. The dataset is divided into a 

training set (1,440 images), validation set (480 images), and testing set (480 images), ensuring a 

comprehensive evaluation of the model's performance. 

Finger Vein Recognition Dataset: The Finger Vein Recognition Dataset consists of 3,000 images 

from 250 subjects, with each subject providing 12 images of their finger veins. Captured using a near-

infrared imaging technique, these images offer clear visibility of the vein patterns. The dataset is 

partitioned into training (1,800 images), validation (600 images), and testing (600 images) sets. This 

partitioning facilitates a robust training regime and an unbiased evaluation of the model's recognition 

capabilities. 

The MDVPR model is implemented using a modified Densenet-161 architecture, with the following 

key parameters set for the experiments: 

• Learning Rate: Initialized at 0.001, with a decay factor of 0.1 applied every 30 epochs to fine-

tune the learning process. 

• Batch Size: Set to 32 for training and validation phases to balance computational efficiency 

and model performance. 

• Epochs: The model is trained for up to 100 epochs, with early stopping implemented based on 

the validation loss to prevent overfitting. 

• Optimizer: Adam optimizer is employed for its adaptive learning rate capabilities, enhancing 

the convergence speed. 

• Loss Function: Cross-entropy loss is utilized for its effectiveness in classification tasks, 

especially with imbalanced datasets & samples. 

The model's performance is assessed using the following metrics: precision, accuracy, recall, delay 

(measured in milliseconds), Area Under the Curve (AUC), and specificity levels. These metrics 

provide a comprehensive overview of the model's efficacy in vein pattern recognition process. 

The experiments are conducted on a system equipped with an NVIDIA RTX 3090 GPU, 32GB RAM, 

and an Intel i9 processor, ensuring high computational power for model training and evaluation. The 

software environment is set up with Python 3.8, PyTorch 1.8 for deep learning model implementation, 

and CUDA 11.2 for GPU accelerations. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

704 https://internationalpubls.com 

Images from both datasets undergo preprocessing steps, including resizing to 224x224 pixels to fit the 

input dimension of the Densenet-161 model, normalization using the mean and standard deviation of 

the ImageNet dataset, and augmentation techniques such as rotation and horizontal flipping to enhance 

model generalizability levels. 

Based on this setup, equations 13, 14, and 15 were used to assess the precision (P), accuracy (A), and 

recall (R), levels based on this technique, while equations 16 & 17 were used to estimate the overall 

precision (AUC) & Specificity (Sp) as follows, 

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃

𝑇𝑃 +  𝐹𝑃
… (13) 

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 +  𝑇𝑁

𝑇𝑃 +  𝑇𝑁 +  𝐹𝑃 +  𝐹𝑁
… (14) 

𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃

𝑇𝑃 +  𝐹𝑁
… (15) 

𝐴𝑈𝐶 = ∫ 𝑇𝑃𝑅(𝐹𝑃𝑅)𝑑𝐹𝑃𝑅 … (16) 

𝑆𝑝 =
𝑇𝑁

𝑇𝑁 + 𝐹𝑃
… (17) 

There are three different kinds of test set predictions: True Positive (TP) (number of events in test sets 

that were correctly predicted as positive), False Positive (FP) (number of instances in test sets that 

were incorrectly predicted as positive), and False Negative (FN) (number of instances in test sets that 

were incorrectly predicted as negative; this includes Normal Instance Samples). The documentation 

for the test sets makes use of all these terminologies. To determine the appropriate TP, TN, FP, and 

FN values for these scenarios, we compared the projected Vien Instances likelihood to the actual Vien 

Instances status in the test dataset samples using the EIFNet [2], NBGAN [4], and FVViT [8] 

techniques. As such, we were able to predict these metrics for the results of the suggested model 

process. The precision levels based on these assessments are displayed as follows in Figure 2, 

 
Figure 2. Observed Precision during Identification for Viens from Clinical Samples 

At lower NTS, such as 63k, MDVPR demonstrates a precision of 91.66%, surpassing EIFNet's 

89.40%, NB-GAN's 86.50%, and FVViT's 89.61%. This trend of superior performance continues as 

the NTS increases, with MDVPR achieving a precision of 93.68% at 117k samples, compared to 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

705 https://internationalpubls.com 

EIFNet's 88.08%, NB-GAN's 91.06%, and FVViT's 90.85%. Such numerical comparisons illustrate 

the enhanced capability of MDVPR in handling varying volumes of data, maintaining high precision 

even as the complexity and quantity of the test samples increase. 

Notably, at the highest volume of test samples, 1,080,000, MDVPR achieves an unprecedented 

precision of 98.58%, significantly outperforming EIFNet's 85.98%, NB-GAN's 85.48%, and FVViT's 

91.96%. This remarkable precision underscores the model's robustness and adaptability to large-scale 

clinical datasets, a critical requirement for deployment in clinical scenarios where the volume of data 

can be vast and varied. 

The reasons behind MDVPR's superior performance can be attributed to several factors. Firstly, the 

modification of the Densenet-161 structure, particularly the removal of the conventional classification 

layer and its transformation into a robust feature embedder, allows for the extraction of highly distinct 

vein features. This strategic alteration is instrumental in capturing the unique identifiers crucial in 

biomedical systems, enabling the model to achieve higher precision by reducing intra-class variations 

while amplifying inter-class differences. 

Furthermore, the methodology employed, involving meticulous training and validation phases 

followed by a testing phase under a verification scenario, ensures that the model is thoroughly 

optimized for vein pattern recognition. The use of a dual-input system for comparing known and 

unknown vein patterns using Euclidean distance quantifies similarity with remarkable accuracy, 

contributing to the model's overall performance. 

The impact of MDVPR's enhanced precision is profound. By significantly reducing error rates and 

susceptibility to environmental variations, the model promises to revolutionize vein detection systems 

used in clinical scenarios. Its application extends beyond mere identification, potentially aiding in 

critical medical procedures such as intravenous catheterization, blood check-ups, and the analysis of 

veins related to various diseases. The model's ability to work efficiently and accurately across different 

scales of data not only demonstrates its technical superiority but also its potential to improve patient 

care and medical diagnostics, making it a valuable contribution to the field of biomedical systems. 

Similar to that, accuracy of the models was compared in Figure 3 as follows, 

 
Figure 3. Observed Accuracy during Identification for Viens from Clinical Samples 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

706 https://internationalpubls.com 

Analyzing the data across various numbers of test samples (NTS), MDVPR consistently demonstrates 

superior accuracy. For instance, at a lower NTS of 63,000, MDVPR achieves an accuracy of 89.15%, 

compared to EIFNet's 87.24%, NB-GAN's 86.21%, and FVViT's 87.91%. This trend of enhanced 

performance is evident across all sample sizes, culminating in an accuracy of 97.57% at 1,080,000 

NTS for MDVPR, significantly higher than the comparisons: EIFNet's 82.85%, NB-GAN's 85.12%, 

and FVViT's 91.70%. 

The superior accuracy of MDVPR can be attributed to its innovative architecture and training 

methodology. The modified Densenet-161 model, with its robust feature embedder, is adept at 

extracting distinctive vein features, which is instrumental in improving the model's ability to recognize 

vein patterns accurately. This capability is crucial for reducing false positives and negatives, thereby 

enhancing the model's reliability and trustworthiness in clinical applications. 

The impact of increased accuracy on clinical scenarios is profound. Higher accuracy ensures that vein 

pattern recognition systems can be effectively used in various critical medical procedures, such as 

intravenous catheterization, where precise vein identification is paramount. Furthermore, accurate vein 

pattern recognition can aid in the early detection of diseases by analyzing blood flow and vein 

anomalies, leading to timely interventions and treatments. 

Moreover, the reliability of vein recognition systems, underscored by high accuracy rates, can 

significantly reduce the procedural time and improve the overall efficiency of clinical operations. This 

is particularly important in high-pressure environments where time is of the essence, and the margin 

for error is minimal. 

The MDVPR model's performance, highlighted by its exceptional accuracy across a broad spectrum 

of test samples, underscores its potential to revolutionize vein recognition in biomedical systems. By 

significantly outperforming existing models, MDVPR offers a promising solution to the challenges 

faced by traditional vein detection systems, including high error rates and susceptibility to 

environmental variations. Its deployment in clinical settings could lead to improved patient outcomes, 

streamlined procedures, and enhanced diagnostic capabilities, marking a significant advancement in 

the field of medical biological s. Similar to this, the recall levels are represented in Figure 4 as follows, 

 
Figure 4. Observed Recall during Identification for Viens from Clinical Samples 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

707 https://internationalpubls.com 

The Modified Densenet-161 Model for Advanced Vein Pattern Recognition in Biomedical Systems 

(MDVPR) consistently exhibits superior recall rates across various test sample sizes (NTS), from 

63,000 to 1,080,000, when compared to EIFNet [2], NB-GAN [4], and FVViT [8]. For instance, at an 

NTS of 63,000, MDVPR achieved a remarkable recall of 96.20%, significantly outperforming EIFNet 

(89.89%), NB-GAN (90.81%), and FVViT (87.33%). This trend of MDVPR achieving higher recall 

rates continues across the board, with notable performance at 153,000 NTS where MDVPR reaches 

97.12% recall, compared to the next highest, FVViT, at 94.72%. 

Such high recall rates are indicative of MDVPR's efficiency in identifying vein patterns accurately, 

minimizing the risk of false negatives. In clinical scenarios, this is particularly impactful as it ensures 

that patients with vein pattern-related conditions are correctly identified and not overlooked. For 

procedures such as intravenous catheterization, where precise vein detection is crucial, the high recall 

rate of MDVPR can significantly reduce complications arising from incorrect vein identification. 

Furthermore, in the context of disease diagnosis and treatment where vein patterns play a role, 

MDVPR's ability to accurately identify relevant vein characteristics can lead to earlier detection of 

conditions, potentially leading to more effective treatment plans and better patient outcomes. The high 

recall rate implies that the system can reliably capture and identify vein patterns, even in challenging 

or low-contrast images, making it a valuable tool in a clinical setting. 

The superior recall rates of MDVPR also suggest its capability to perform well under various 

conditions and with different patient demographics, which is essential for the diverse range of clinical 

scenarios encountered in practice. By ensuring a low rate of missed positive identifications, MDVPR 

can contribute to enhanced patient safety, reduce diagnostic errors, and improve the overall efficiency 

of clinical procedures. 

In conclusion, the high recall observed with the MDVPR model signifies a substantial advancement in 

vein pattern recognition technology, offering a promising solution to the challenges faced by existing 

models. Its deployment in clinical settings could lead to significant improvements in patient care, 

diagnostic accuracy, and procedural efficiency, highlighting the model's potential to impact positively 

on the medical field sets. Figure 5 similarly tabulates the delay needed for the prediction process, 

 
Figure 5. Observed Delay during Identification for Viens from Clinical Samples 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

708 https://internationalpubls.com 

At an NTS of 63,000, MDVPR demonstrates a remarkable processing speed with an observed delay 

of only 99.30ms, which is faster compared to EIFNet's 126.92ms, NB-GAN's 111.67ms, and even 

FVViT's 101.32ms. This trend of MDVPR exhibiting lower or competitive processing times continues 

across various test sample sizes, highlighting its efficiency in quickly processing and analyzing images 

for vein pattern recognition. 

For example, at higher NTS counts like 900,000 and 1,080,000, MDVPR maintains a delay of 

102.15ms and 108.98ms, respectively, showcasing its scalability and capability to handle large datasets 

without significant increases in processing time. This is particularly notable when compared to other 

models, where delays can exceed 130ms in some cases, as observed with EIFNet and NB-GAN. 

The impacts of these observed delays in clinical scenarios are significant. In medical settings, where 

time efficiency can directly influence patient throughput and the quality of care, the reduced processing 

time of MDVPR can lead to faster diagnosis and treatment initiation. For procedures requiring real-

time or near-real-time data processing, such as intravenous catheterization or emergency diagnostics, 

the efficiency of MDVPR can enhance clinical workflows, reduce patient wait times, and potentially 

decrease the stress and discomfort experienced by patients during vein detection processes. 

Furthermore, the ability of MDVPR to maintain low processing times across large datasets suggests 

its applicability in high-volume clinical environments, such as large hospitals or clinics conducting 

numerous vein pattern recognition tasks daily for different use cases. This efficiency can contribute to 

more streamlined operations, allowing healthcare providers to allocate resources more effectively and 

focus on delivering higher-quality patient care scenarios. 

In summary, the MDVPR model not only excels in accuracy and recall but also stands out for its 

processing efficiency, making it a highly desirable solution in the fast-paced clinical setting. Its 

adoption could significantly impact the efficiency of medical procedures that rely on vein pattern 

recognition, underscoring the model's potential to improve both the patient experience and the overall 

effectiveness of healthcare delivery systems. Similarly, the AUC levels can be observed from figure 6 

as follows, 

 
Figure 6. Observed AUC during Identification for Viens from Clinical Samples 

At a lower number of test samples (NTS) of 63,000, MDVPR achieves an AUC of 85.97%, 

significantly outperforming EIFNet (78.77%), NB-GAN (77.50%), and FVViT (71.98%). This trend 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

709 https://internationalpubls.com 

of MDVPR maintaining a higher AUC is consistent across different NTS, showcasing its robustness 

in accurately classifying vein patterns across a wide range of clinical scenarios. Particularly 

noteworthy is the AUC score of 94.72% at 153,000 NTS, which far surpasses the scores of the other 

models, underscoring MDVPR's exceptional performance. 

In clinical settings, the high AUC achieved by MDVPR has profound implications. A model with a 

high AUC can reliably differentiate between the presence and absence of specific vein patterns, which 

is crucial for diagnosing vascular diseases, planning intravenous catheterization, and other medical 

procedures where vein detection plays a critical role. The ability to accurately classify and identify 

vein patterns reduces the likelihood of false positives or negatives, which in turn minimizes the risk of 

incorrect diagnoses or unnecessary interventions. 

Moreover, the high discriminative power of MDVPR ensures that the system is less likely to be 

influenced by variations in image quality, patient movement, or other environmental factors that 

typically challenge vein pattern recognition systems. This resilience enhances the model's applicability 

in a wide range of clinical environments, from high-throughput settings like busy hospitals to more 

controlled environments like diagnostic laboratories. 

Additionally, the confidence in MDVPR's ability to accurately identify vein patterns can lead to greater 

trust in automated systems by healthcare professionals. This trust is crucial for the adoption of such 

technologies in clinical practice, where the stakes are high, and the accuracy of diagnostics and 

treatments can significantly impact patient outcomes. 

In summary, the superior AUC scores of MDVPR demonstrate its effectiveness and reliability in vein 

pattern recognition. Its deployment in clinical scenarios could lead to improvements in patient care 

through more accurate diagnostics, reduced procedural times, and enhanced overall healthcare 

delivery. The model's high level of accuracy and reliability, as evidenced by its AUC scores, marks a 

significant advancement in biomedical imaging and has the potential to transform the landscape of 

medical diagnostics and treatment planning process. Similarly, the Specificity levels can be observed 

from figure 7 as follows, 

 
Figure 7. Observed Specificity during Identification for Viens from Clinical Samples 

In the data provided, the Modified Densenet-161 Model for Advanced Vein Pattern Recognition in 

Biomedical Systems (MDVPR) consistently shows high specificity across a range of test sample sizes 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

710 https://internationalpubls.com 

(NTS), from 63,000 to 1,080,000, when compared to EIFNet [2], NB-GAN [4], and FVViT [8]. For 

instance, at the lower NTS of 63,000, MDVPR achieves a specificity of 86.06%, which is noticeably 

higher than EIFNet's 78.29%, NB-GAN's 77.56%, and FVViT's 82.12%. This trend of MDVPR 

exhibiting superior specificity continues throughout the various NTS levels, reaching up to 94.85% at 

1,080,000 NTS, showcasing its exceptional ability to accurately identify non-vein patterns or negative 

cases. 

In clinical settings, high specificity has profound impacts. It ensures that patients are not misdiagnosed 

with conditions they do not have, thereby avoiding unnecessary anxiety, treatment, or intervention. 

For example, in procedures where vein identification is crucial for delivering treatments (such as 

targeted drug delivery), high specificity ensures that the system accurately identifies the targeted veins 

and avoids incorrect areas, thereby minimizing potential side effects or complications. 

Moreover, the high specificity of MDVPR can contribute to the efficiency of clinical workflows. By 

reducing the number of false positives, the system can help streamline patient care processes, allowing 

healthcare professionals to focus on patients who require intervention. This can be particularly 

beneficial in high-volume clinical environments where time and accuracy are paramount, such as 

emergency departments or outpatient clinics. 

Additionally, the confidence in negative results provided by a system with high specificity can be 

crucial in monitoring and screening contexts. For instance, in the screening for vascular diseases or 

conditions that manifest through changes in vein patterns, high specificity ensures that individuals 

without these conditions are accurately identified, allowing for more focused and resource-efficient 

care for those who do test positive in different use case scenarios. 

In summary, the superior specificity observed with the MDVPR model highlights its potential as a 

reliable and efficient tool in vein pattern recognition for biomedical applications. Its ability to 

accurately rule out negative cases can lead to significant improvements in patient care, reduce 

unnecessary medical interventions, and enhance the overall efficiency of clinical operations, marking 

a significant step forward in the application of artificial intelligence in healthcare scenarios. 

5.  Conclusion and Future Scope 

The exploration and development of the Modified Densenet-161 Model for Advanced Vein Pattern 

Recognition in Biomedical Systems (MDVPR) represent a significant leap forward in the domain of 

biological identification and medical diagnostics. This study's findings illuminate the model's 

unparalleled precision, accuracy, recall, and specificity in identifying vein patterns from clinical 

samples, alongside its expedited processing capabilities as evidenced by reduced observed delays. 

Notably, the MDVPR model achieves a marked improvement over existing methodologies, 

particularly in the realms of precision and AUC metrics across a variety of test sample sizes, 

underscoring its robustness and adaptability to diverse clinical scenarios. 

The implications of this work are profound, extending beyond the immediate sphere of biological 

authentication to enrich medical procedures where vein pattern recognition plays a pivotal role. The 

enhanced performance of the MDVPR model, characterized by its high accuracy and low delay, offers 

the potential to revolutionize patient care through improved efficiency and reliability of vein detection 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

711 https://internationalpubls.com 

in procedures such as intravenous catheterization and the early diagnosis of vascular diseases. 

Moreover, the model's high specificity and recall rates minimize the likelihood of false positives and 

negatives, thereby bolstering the confidence of healthcare professionals in the deployment of 

automated vein recognition technologies. 

Looking ahead, the scope for further refinement and application of the MDVPR model is vast. Future 

investigations could explore the integration of additional modalities, such as thermal imaging or pulse 

oximetry data, to further enhance the model's sensitivity and specificity. Such multidimensional data 

incorporation could unlock new pathways for detecting subtle vein pattern changes, offering earlier 

diagnostic cues for diseases like deep vein thrombosis or varicose veins. Moreover, the scalability and 

efficiency of the MDVPR model beckon its application in large-scale health monitoring systems, 

potentially enabling the development of wearable technologies for continuous vascular health 

assessment. 

The adaptability of the model to contactless imaging techniques also opens up exciting avenues for its 

application in settings where hygiene and minimal patient contact are paramount, such as in infectious 

disease wards or during pandemic scenarios. Furthermore, the potential for cross-disciplinary 

applications, including secure access control in sensitive environments based on biological 

authentication, highlights the versatility of the MDVPR model process. 

Thus, the MDVPR model stands as a testament to the potential of advanced machine learning 

techniques to address and surpass the limitations of existing vein pattern recognition methodologies. 

Its development not only advances the state of the art in biological authentication but also opens new 

horizons for enhancing patient care through more accurate, efficient, and reliable medical diagnostics. 

As this field continues to evolve, the MDVPR model's foundational contributions will undoubtedly 

inspire further innovation, driving the frontier of biological and medical technology towards new 

heights of excellence in clinical use cases. 

Financing and Declaration of Conflict of Interests: 

The authors don’t have any conflict of interest among them. The authors certify that they have NO 

affiliations with or involvement in any organization or entity with any financial interest or non-

financial interest (such as personal or professional relationships, affiliations, knowledge, or beliefs) in 

the subject matter or materials discussed in this manuscript. 

Ethical Approval : This article does not contain any studies with human participants or animals 

performed by any of the authors. 

6. References 

[1] A. Krishnan and T. Thomas, "Finger Vein Recognition Based on Anatomical Features of Vein Patterns," in IEEE 

Access, vol. 11, pp. 39373-39384, 2023, doi: 10.1109/ACCESS.2023.3253203. 

keywords: {Biomedicals (access control);Feature extraction;Fingers;Pattern recognition;Pattern matching;Image 

recognition;Veins;Authentication;FEBA classification;F2EB2A;feature representation;finger vein 

biomedicals;matching;vein anatomy}, 

[2] Y. Song, P. Zhao, W. Yang, Q. Liao and J. Zhou, "EIFNet: An Explicit and Implicit Feature Fusion Network for 

Finger Vein Verification," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 5, pp. 

2520-2532, May 2023, doi: 10.1109/TCSVT.2022.3224203. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

712 https://internationalpubls.com 

keywords: {Veins;Feature extraction;Fingers;Convolution;Deep learning;Data mining;Image recognition;Finger vein 

verification;deep learning;feature extraction;vein pattern extraction;feature fusion}, 

[3] G. Xu et al., "Portal Vein Pressure Estimation and Portal Hypertension Discrimination Based on Subharmonic 

Scattering of Ultrasound Contrast Agent Microbubbles," in IEEE Transactions on Biomedical Engineering, vol. 71, 

no. 1, pp. 150-159, Jan. 2024, doi: 10.1109/TBME.2023.3293952. 

keywords: {Acoustics;Portals;Liver;Ultrasonic imaging;Biomedical imaging;Veins;Pressure measurement;Portal 

hypertension;portal vein pressure;noninvasive blood pressure estimation;subharmonic scattering;ultrasound contrast 

agent microbubbles}, 

[4] J. He et al., "Finger Vein Image Deblurring Using Neighbors-Based Binary-GAN (NB-GAN)," in IEEE Transactions 

on Emerging Topics in Computational Intelligence, vol. 7, no. 2, pp. 295-307, April 2023, doi: 

10.1109/TETCI.2021.3097734. 

keywords: {Veins;Kernel;Mathematical models;Image restoration;Generators;Convolution;Training;Finger 

vein;GAN;image deblurring;texture loss}, 

[5] L. Yang, X. Liu, G. Yang, J. Wang and Y. Yin, "Small-Area Finger Vein Recognition," in IEEE Transactions on 

Information Forensics and Security, vol. 18, pp. 1914-1925, 2023, doi: 10.1109/TIFS.2023.3258252. 

keywords: {Veins;Image recognition;Feature extraction;Databases;Convolutional neural networks;Dictionaries;Learning 

systems;Finger vein recognition;small-area image;embedded sensors;dictionary learning}, 

[6] P. Zhao et al., "Single-Sample Finger Vein Recognition via Competitive and Progressive Sparse Representation," in 

IEEE Transactions on Biomedicals, Behavior, and Identity Science, vol. 5, no. 2, pp. 209-220, April 2023, doi: 

10.1109/TBIOM.2022.3226270. 

keywords: {Veins;Fingers;Training;Dictionaries;Biomedicals (access control);Lighting;Indexes;Finger vein;finger dorsal 

texture;progressive sparse representation;competitive sparse fusion}, 

[7] Z. Huang and C. Guo, "Toward Cross-Dataset Finger Vein Recognition With Single-Source Data," in IEEE 

Transactions on Instrumentation and Measurement, vol. 73, pp. 1-12, 2024, Art no. 2506012, doi: 

10.1109/TIM.2023.3348905. 

keywords: {Data models;Training;Adaptation models;Task analysis;Fingers;Image recognition;Veins;Cross-dataset 

recognition;domain adaptation;domain alignment;fast transfer learning;finger vein recognition (FVR)}, 

[8] X. Li and B. -B. Zhang, "FV-ViT: Vision Transformer for Finger Vein Recognition," in IEEE Access, vol. 11, pp. 

75451-75461, 2023, doi: 10.1109/ACCESS.2023.3297212. 

keywords: {Transformers;Feature extraction;Computational modeling;Databases;Computer architecture;Computer 

vision;Training data;Biomedicals (access control);Veins;Vision transformer;finger vein recognition;computer 

vision}, 

[9] G. Tong, H. Jiang, T. Shi, X. -H. Han and Y. -D. Yao, "A Lightweight Network for Contextual and Morphological 

Awareness for Hepatic Vein Segmentation," in IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 10, 

pp. 4878-4889, Oct. 2023, doi: 10.1109/JBHI.2023.3305644. 

keywords: {Veins;Liver;Three-dimensional displays;Morphology;Feature extraction;Image segmentation;Task 

analysis;Medical image segmentation;Context awareness;Hepatic veins;Attention mechanism;lightweight}, 

[10] Y. Huang, H. Ma and M. Wang, "Axially Enhanced Local Attention Network for Finger Vein Recognition," in IEEE 

Transactions on Instrumentation and Measurement, vol. 72, pp. 1-10, 2023, Art no. 5020210, doi: 

10.1109/TIM.2023.3291785. 

keywords: {Feature extraction;Veins;Fingers;Data mining;Image recognition;Topology;Network topology;Attentional 

mechanisms;deep learning;feature extraction;finger vein recognition;lightweight network}, 

[11] C. -G. Ri, K. -I. Ri and J. -H. Ri, "Notice of Removal: Robust and Efficient Finger Vein Recognition Using Binarized 

Difference of Gabor Jet on Grid," in IEEE Transactions on Information Forensics and Security, vol. 18, pp. 177-189, 

2023, doi: 10.1109/TIFS.2022.3217396. 

[12] J. Huang, A. Zheng, M. S. Shakeel, W. Yang and W. Kang, "FVFSNet: Frequency-Spatial Coupling Network for 

Finger Vein Authentication," in IEEE Transactions on Information Forensics and Security, vol. 18, pp. 1322-1334, 

2023, doi: 10.1109/TIFS.2023.3238546. 

keywords: {Veins;Fingers;Feature extraction;Frequency-domain analysis;Authentication;Band-pass filters;Low-pass 

filters;Finger vein;biomedicals;authentication;computer vision;neural network}, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

713 https://internationalpubls.com 

[13] J. Chang, T. Lai, L. Yang, C. Fang, Z. Li and H. Fujita, "FV-DGNN: A Distance-Based Graph Neural Network for 

Finger Vein Recognition," in IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1-11, 2023, Art 

no. 2523011, doi: 10.1109/TIM.2023.3301062. 

keywords: {Feature extraction;Graph neural networks;Data mining;Computer architecture;Computational 

modeling;Databases;Convolutional neural networks;Convolutional autoencoder (CAE) architecture;depthwise 

separable convolution layer;distance distribution;finger vein recognition;graph neural network (GNN)}, 

[14] H. Qin, R. Hu, M. A. El-Yacoubi, Y. Li and X. Gao, "Local Attention Transformer-Based Full-View Finger-Vein 

Identification," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 6, pp. 2767-2782, 

June 2023, doi: 10.1109/TCSVT.2022.3227385. 

keywords: {Veins;Fingers;Cameras;Three-dimensional displays;Feature extraction;Face recognition;Deep learning;Multi-

view vein identification;full-view finger-vein images;transformer;deep learning}, 

[15] H. Qin, C. Fan, S. Deng, Y. Li, M. A. El-Yacoubi and G. Zhou, "AG-NAS: An Attention GRU-Based Neural 

Architecture Search for Finger-Vein Recognition," in IEEE Transactions on Information Forensics and Security, vol. 

19, pp. 1699-1713, 2024, doi: 10.1109/TIFS.2023.3340915. 

keywords: {Computer architecture;Task analysis;Neural networks;Veins;Convolutional neural networks;Logic 

gates;Fingerprint recognition;Finger-vein recognition;deep learning;neural architecture search (NAS);gated recurrent 

unit (GRU);self-attention}, 

[16] S. Li, R. Ma, J. Zhou, B. Zhang and L. Wu, "Joint Discriminative Analysis With Low-Rank Projection for Finger 

Vein Feature Extraction," in IEEE Transactions on Information Forensics and Security, vol. 19, pp. 959-969, 2024, 

doi: 10.1109/TIFS.2023.3326364. 

keywords: {Feature extraction;Sparse matrices;Image recognition;Biomedicals (access control);Adaptation 

models;Databases;Data models;Finger vein recognition;jointly embedding;discriminative analysis;low-rank 

representation;domain adaptation}, 

[17] Z. Zhang, G. Chen, W. Zhang and H. Wang, "Finger Vein Recognition Based on ResNet With Self-Attention," in 

IEEE Access, vol. 12, pp. 1943-1951, 2024, doi: 10.1109/ACCESS.2023.3347922. 

keywords: {Convolutional neural networks;Feature extraction;Fingers;Transformers;Analytical models;Iris 

recognition;Finger vein recognition;deep learning;ResNet;self-attention;variable learning rate}, 

[18] R. Garcia-Martin and R. Sanchez-Reillo, "Vision Transformers for Vein Biomedical Recognition," in IEEE Access, 

vol. 11, pp. 22060-22080, 2023, doi: 10.1109/ACCESS.2023.3252009. 

keywords: {Machine learning;Transformers;Biomedicals (access control);Databases;Image recognition;Convolutional 

neural networks;Transfer learning;Vision transformers;vein biomedical recognition;deep learning;convolutional 

neural networks;finger veins;transfer learning;machine learning;artificial intelligence;biomedicals on mobile 

devices;contactless wrist vascular database;hand palm identification}, 

[19] Z. Pan, J. Wang, Z. Shen and S. Han, "Disentangled Representation and Enhancement Network for Vein 

Recognition," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 8, pp. 4164-4176, 

Aug. 2023, doi: 10.1109/TCSVT.2023.3241054. 

keywords: {Veins;Feature extraction;Shape;Image recognition;Task analysis;Data mining;Training;Vein 

recognition;disentangled representation and enhancement network (DRE-Net);multi-scale attention residual block 

(MSARB);weight-guided feature enhancement module (WGFEM)}, 

[20] R. Güldenring, R. E. Andersen and L. Nalpantidis, "Zoom in on the Plant: Fine-Grained Analysis of Leaf, Stem, and 

Vein Instances," in IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1588-1595, Feb. 2024, doi: 

10.1109/LRA.2023.3346807. 

keywords: {Robots;Measurement;Annotations;Task analysis;Grasslands;Feature extraction;Crops;Robotics and 

automation in agriculture and forestry;field robots;phenotyping;keypoint-guided polylines;image dataset}, 

[21] Y. Ma, H. Huang, D. Luo, S. Zhang, W. Kang and D. Xie, "Focal Contrastive Learning for Palm Vein Authentication," 

in IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1-15, 2023, Art no. 5025315, doi: 

10.1109/TIM.2023.3304689. 

keywords: {Authentication;Feature extraction;Task analysis;Training;Deep learning;Computational modeling;Loss 

measurement;Authentication;biomedical technology;contrastive learning;hard sample mining;palm vein}, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 8s (2024) 

 

714 https://internationalpubls.com 

[22] Y. Li, S. Ruan, H. Qin, S. Deng and M. A. El-Yacoubi, "Transformer Based Defense GAN Against Palm-Vein 

Adversarial Attacks," in IEEE Transactions on Information Forensics and Security, vol. 18, pp. 1509-1523, 2023, 

doi: 10.1109/TIFS.2023.3243782. 

keywords: {Veins;Feature extraction;Deep learning;Convolutional neural networks;Image recognition;Perturbation 

methods;Shape;Local transformer;defense GAN;purifier;palm-vein recognition;adversarial attacks}, 

[23] H. Qin, C. Gong, Y. Li, X. Gao and M. A. El-Yacoubi, "Label Enhancement-Based Multiscale Transformer for Palm-

Vein Recognition," in IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1-17, 2023, Art no. 

2509217, doi: 10.1109/TIM.2023.3261909. 

keywords: {Veins;Feature extraction;Convolutional neural networks;Face recognition;Deep 

learning;Databases;Transformers;Deep learning;graph convolutional network (GCN);label enhancement (LE);palm-

vein identification;transformer}, 

[24] M. Zhang, J. Li, Y. Wang and G. Xu, "CECNet: Coordinate Encoding Competitive Neural Network For Palm Vein 

Recognition by Soft Large Margin Centralized Cosine Loss," in IEEE Access, vol. 11, pp. 141329-141342, 2023, doi: 

10.1109/ACCESS.2023.3341229. 

keywords: {Feature extraction;Gabor filters;Encoding;Convolutional neural networks;Information filters;Adaptive 

filters;Support vector machines;Palm vein recognition;coordinate encoding competitive neural network;soft large 

margin centralized cosine loss;adaptive Gabor filter encoders;feature extraction}, 

[25] T. -V. Nguyen, S. -J. Horng, D. -T. Vu, H. Chen and T. Li, "LAWNet: A Lightweight Attention-Based Deep Learning 

Model for Wrist Vein Verification in Smartphones Using RGB Images," in IEEE Transactions on Instrumentation 

and Measurement, vol. 72, pp. 1-10, 2023, Art no. 2531810, doi: 10.1109/TIM.2023.3328702. 

keywords: {Wrist;Feature extraction;Smart phones;Biological system modeling;Image color analysis;Sensors;Spatial 

databases;Attention mechanism;deep learning;saturation channel;wrist vein extraction;wrist vein recognition}, 

[26] A. Azapagic et al., "A Novel Vascular Anastomotic Coupling Device for End-to-End Anastomosis of Arteries and 

Veins," in IEEE Transactions on Biomedical Engineering, vol. 71, no. 2, pp. 542-552, Feb. 2024, doi: 

10.1109/TBME.2023.3308890. 

keywords: {Surgery;Couplers;Thrombosis;Arteries;Veins;Couplings;Microsurgery;Arterial coupler;end-to-end 

anastomosis;free tissue transfer;microsurgery;microvascular anastomotic coupler (MAC);venous coupler}, 

[27] T. Eglitis, E. Maiorana and P. Campisi, "Open-Source Finger Vein Acquisition Device for Biomedical Applications," 

in IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1-12, 2023, Art no. 5031912, doi: 

10.1109/TIM.2023.3324681. 

keywords: {Fingers;Veins;Databases;Biomedicals (access control);Lighting;Absorption;Software;Biomedicals;finger 

vein;open source;prototype;vascular biomedicals}, 

[28] R. S. Kuzu, E. Maiorana and P. Campisi, "Gender-Specific Characteristics for Hand-Vein Biomedical Recognition: 

Analysis and Exploitation," in IEEE Access, vol. 11, pp. 11700-11710, 2023, doi: 10.1109/ACCESS.2023.3239894. 

keywords: {Gender issues;Biomedicals (access control);Pattern recognition;Feature extraction;Character recognition;Face 

recognition;Databases;Deep learning;Biomedical recognition;gender recognition;vein biomedicals;deep learning}, 

[29] S. Li, B. Zhang, L. Wu, R. Ma and X. Ning, "Robust and Sparse Least Square Regression for Finger Vein and Finger 

Knuckle Print Recognition," in IEEE Transactions on Information Forensics and Security, vol. 19, pp. 2709-2719, 

2024, doi: 10.1109/TIFS.2024.3352389. 

keywords: {Feature extraction;Matrix decomposition;Sparse matrices;Biomedicals (access control);Representation 

learning;Matrix converters;Image recognition;Finger-based biomedicals;projection learning;sparse transformation 

matrix;least square regression (LSR)}, 

[30] O. C. A. Gadea and S. D. Khan, "Detection of Bastnäsite-Rich Veins in Rare Earth Element Ores Through 

Hyperspectral Imaging," in IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1-4, 2023, Art no. 5502204, 

doi: 10.1109/LGRS.2023.3249624. 

keywords: {Reflectivity;Hyperspectral imaging;Rocks;Indexes;Absorption;Image color analysis;Cameras;Hyperspectral 

imaging;mining industry;rare earth metals;spectroradiometers}.