DermaNetX: An Efficient Deep Learning Framework For Automated Skin Disease Classification

Abstract

Deep learning-based automated skin disease classification has become a vital tool to facilitate early clinical diagnosis and lessen reliance on the availability of specialists. This paper introduces DermaNetX, a novel hybrid convolutional neural network framework for multi-class dermoscopic image classification that combines ResNet-50 and EfficientNet-B4 via a dual-branch transfer learning architecture. DermaNetX uses a twophase training approach: frozen backbone pre-training and selective fine-tuning of the final two convolutional blocks of each backbone. It was trained and assessed on the HAM10000 dataset, which includes 10,015 dermoscopic images from seven different disease categories. The framework incorporates a WeightedRandomSampler an extensive augmentation pipeline, and weighted cross-entropy loss to address severe class imbalance. The fused 3840-dimensional feature representation is passed through a three-layer classification head with batch normalisation and dropout regularisation. On the held-out test set, DermaNetX achieves a macro-averaged AUC of 0.9518 and a balanced accuracy of 79.45%, with dermatofibroma recall reaching 94% and melanoma recall reaching 80%. Clinicians can confirm the anatomical focus of each prediction thanks to Grad-CAM's spatial explainability. Ablation verifies that the hybrid architecture performs better than single-backbone baselines in every evaluation metric.