Risk Adaptive Triple Stream Vision Transformer for Dermatological Triage

Sivadasan Achary, Saran Das (2025) Risk Adaptive Triple Stream Vision Transformer for Dermatological Triage. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (731kB) | Preview
Preview	PDF (Configuration Manual) Download (456kB) | Preview

Abstract

Skin diseases are a health care challenge especially in areas where dermatologists are not readily accessible. As much as computer-assisted triage systems hold promise in ranking high-risk cases, the existing approaches currently have pre-trained model bias and are missing such critically important clinical characteristics as calibration and interpretability. Our suggested model is a Risk-Adaptive Triple-Stream ViTra (TS-ViT) which trains on a dataset of dermatological images to identify lesions into the categories of HIGH, MODERATE, and LOW clinical urgency. Its architecture combines three streams, namely global classification to obtain whole-view risk assessment, spatial attention to obtain interpretable heatmaps and uncertainty quantification with Monte Carlo dropout to estimate confidence. TS-ViT (12,345 dermatoscopic images, 40 subclasses) demonstrated accuracy (89.3%) and macro F1-score (0.847), which is significantly less than ViT-B/16 (67.2%), ResNet-50 (63.8%), and MobileNetV3-Large (59.4%). The model had better calibration (ECE: 0.074, Brier Score: 0.098), and good uncertainty-error correlation (Spearman r = 0.68, p < 0.001), which allowed the use of triage policies based on confidence. There was correspondence between the spatial attention maps and the diagnostic criteria, with irregular borders and asymmetric pigmentation being significantly observed in the malignant lesions. Such findings indicate that multi-stream architectures that are cautious of risks are superior to classical accuracy-based models in clinical utility, which is the base of safe and interpretable dermatological triage in constraint resources.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Staikopoulos, Athanasios UNSPECIFIED
Uncontrolled Keywords:	Triple-Stream Vision Transformer; Skin Lesion Classification; Interpretable Deep Learning; Dermatological Triage; Uncertainty Quantification; Model Calibration
Subjects:	R Medicine > RL Dermatology Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision R Medicine > Healthcare Industry Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	09 Sep 2026 09:41
Last Modified:	09 Sep 2026 09:41
URI:	https://norma.ncirl.ie/id/eprint/9914

Actions (login required)

View Item