A Multi-Task Deep Learning Framework for Skin Lesion Classification, ABCDE Feature Quantification, and Evolution Simulation

Harsha Kotla

Keywords: melanoma, skin lesion, ABCDE criteria, deep learning, HAM10000 dataset, Generative adversarial networks (GANs), benign nevi, malignant melanoma

TL;DR: This paper introduces a multi-task deep learning framework that classifies skin lesions and quantifies the ABCDE clinical features to provide interpretable, transparent melanoma diagnosis.

Abstract: Early detection of melanoma significantly improves survival rates, but many deep learning approaches do not justify their predictions with established dermatological assessment metrics. This work introduces a multi-task neural network that classifies skin lesions and quantifies interpretable ABCDE (Asymmetry, Border irregularity, Color variation, Diameter, Evolving) features. Trained on the HAM10000 dataset, the model achieves 89\% accuracy overall and an AUC of 0.96 for the detection of melanoma in addition to providing quantitative scores for each characteristic. In addition, a module for lesion evolution visualizes a simulated ABCD feature trajectory and gives a more interpretable progression pattern from benign to malignant. Because HAM10000 contains only static images, the “E” (Evolving) feature was simulated computationally as it modeled the temporal trajectories of ABCD features in latent space. This improves diagnostic transparency and can assist dermatologists and educators by linking deep learning outputs to established clinical assessment criteria.

Submission Number: 17