Yuvaraj S.

Yuvaraj S.

Auto-generated profile, based on publicly available information

Yuvaraj S.

Department of Computer Science and Engineering, Sri Eshwar College of Engineering, Coimbatore, Tamil Nadu, India

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
26 November 2025 in Journal of Electronics, Electromedical Engineering, and Medical Informatics
Verified
Journal Article
Open access
KVKavithamani V
Kavithamani V
VKV. Kavya
V. Kavya
RSR. Suganthi
R. Suganthi
YSYuvaraj S.
Yuvaraj S.
PMP. Monisha
P. Monisha
APArun Patrick
Arun Patrick
Lung cancer segmentation and classification from computed tomography (CT) images play a vital role in early diagnosis, prognosis assessment, and effective treatment planning. Despite significant progress in medical image analysis, accurate lung lesion analysis remains highly challenging due to overlapping anatomical structures, heterogeneous tissue intensity distributions, irregular and complex tumor shapes, and poorly defined lesion boundaries. These factors often limit the reliability and generalization capability of conventional deep learning models when applied to real-world clinical data. To address these challenges, this paper proposes a Hybrid Swarm-Driven Vision Transformer (HSViT) framework that synergistically combines swarm intelligence with transformer-based deep learning. The processing pipeline begins with Contrast Limited Adaptive Histogram Equalization (CLAHE), which enhances local contrast while suppressing noise amplification, thereby improving the visibility of subtle pulmonary nodules and lesion regions. Subsequently, a U-Net segmentation model optimized using the Coyote Optimization Algorithm (COA) is employed to accurately delineate lung lesions. COA, a swarm-based metaheuristic, adaptively fine-tunes U-Net parameters, enabling improved convergence and more precise boundary detection compared to gradient-based optimization alone. Following segmentation, discriminative lesion features are extracted and passed to the HSViT classifier. The proposed classifier integrates a Dual-Stage Attention Fusion (DSAF) mechanism, which effectively captures both fine-grained local spatial features and long-range global contextual dependencies. The framework achieves a Dice Coefficient of 0.95, an overall classification accuracy of 98.7%, and a minimized training loss of 0.04. These results highlight the strong potential of HSViT for reliable automated lung cancer diagnosis and for supporting clinical decision-making systems in real-world healthcare environments.

26 November 2025 in Journal of Electronics, Electromedical Engineering, and Medical Informatics
Verified
Journal Article
Open access
BJB. K. Jaleesha
B. K. Jaleesha
SSSuganthi Suganthi
Suganthi Suganthi
NPN. K. Priyadharsini
N. K. Priyadharsini
YSYuvaraj S.
Yuvaraj S.
MPM. Pyingkodi
M. Pyingkodi
MVM. Vallikkannu
M. Vallikkannu
Early and reliable identification of diabetic skin complications, including ischemia and infection, is essential for timely clinical intervention and prevention of severe outcomes. Nevertheless, traditional deep learning models often exhibit limited generalization capability and high computational demands, particularly when distinguishing between visually subtle infection types. To overcome these challenges, this study introduces an end-to-end deep learning architecture termed the Enhanced Multi-Resolution Multi-Path Attention Network (EMRMP-Net), specifically designed for robust diabetic lesion classification. A key contribution of this work is the introduction of a trainable attention-based fusion mechanism that adaptively learns to weight and integrate multi-resolution feature maps, enhancing contextual understanding and discriminative performance. To address the prevalent issue of class imbalance in medical imaging datasets, EMRMP-Net utilizes focal loss and domain-tailored data augmentation, thereby promoting stable learning and improved representation of minority classes. Additionally, a shared classification head across multiple resolution pathways enables joint feature optimization, reducing computational redundancy and improving learning efficiency compared to traditional MRMP models. Comprehensive experiments on the publicly available Diabetic Foot Ulcer (DFU) dataset demonstrate that EMRMP-Net surpasses existing state-of-the-art-methods, achieving 98.12% accuracy and 98.14% F1-score for ischemia detection, and 95.27% accuracy with 93.68% F1-score for infection classification. Overall, EMRMP-Net provides a highly effective, computationally efficient, and generalizable framework for automated diabetic skin lesion analysis, demonstrating strong potential for real-world clinical applications. EMRMP-Net is designed as a general framework for diabetic skin lesion analysis, capable of handling diverse lesion characteristics through multi-resolution and attention-based feature learning. However, in this work, the model is explicitly formulated, trained, and evaluated for the clinically critical binary classification task of distinguishing ischemic ulcers from infected ulcers within DFU imagery.

1 September 2026 in bioRxiv
Verified
Preprint
Open access
PDF file
SVSam Victor
Sam Victor
SSSiddharth Sadasivam
Siddharth Sadasivam
GSGugan Satheesh
Gugan Satheesh
SPShravan Aarya Prabhakaran
Shravan Aarya Prabhakaran
SYS Yuvaraj
S Yuvaraj
RSR Subha
R Subha
Marker-gene amplicon workflows run on shared computing infrastructure, where cores, memory and wall time must be requested before a job starts. Those requests are set by convention, and the convention assumes a pipeline scales with the size of its data. For one of the two dominant stages, it does not. We measured a QIIME 2 16S rRNA workflow across nine upper-respiratory samples from a paediatric otitis media cohort. Over a 27-fold range of sequencing depth, DADA2 denoising rose 12.8-fold in wall time while Naïve Bayes classification changed by one per cent across samples of low to moderate richness, 5 to 70 ASVs each and 218 across the pooled cohort. Richness did not saturate, so the invariance is not saturation: classification is dominated by a fixed cost paid before any sequence is examined. Profiling that cost shows it is compute-bound rather than storage-bound, and nearly half of it is the framework importing plugins the job never calls: classifying in an environment holding only the plugins the step needs is 43.6% faster, uses 28% less memory and returns a byte-identical result. Parallelism returns little for denoising and makes classification slower, though running samples concurrently raises throughput 21-fold against one job running alone. Sequences and taxonomy are unchanged across thread counts, so the implied savings cost nothing scientifically. The magnitude of the fixed cost is a property of the installation and the reference artifact rather than a constant of the method, so the protocol, measurements and code are released for the constants to be remeasured where they are to be relied on.