An Ensemble Deep Learning Framework for Automated Multi Class Skin Lesion Classification Using ConvNeXt-Tiny, EfficientNetV2-S, and MobileNetV3

Skin cancer is one of the fastest rising malignancies in the world, where early and accurate diagnosis plays a decisive role in the survival of patients. Even though the performance of convolutional neural networks (CNNs) in dermoscopic image analysis has been impressive, three main issues that hinder their application in the clinical setting remain: severe imbalance in classes, inter-class visual similarity, overfitting, and lack of interpretability. In this paper, a three-stream ensemble deep learning model that combines ConvNeXt-Tiny, EfficientNetV2-S, and MobileNetV3-Large has been proposed to classify automated multiclass skin lesion on the HAM10000 dataset. Balanced stratified sampling strategy (BalancedDataGen) is used in order to take care of class imbalance without synthetic oversampling in order to have equal contribution of classes in the training. More successful feature diversity and generalization are further improved with the help of adversarial Albumentations-based data augmentation. Each backbone model is fine-tuned separately based on label smoothing, dropout regularization, early stopping, and adaptive learning rate scheduling. The probability-level ensemble averaging is done to give final predictions. The post-hoc explainability of the results is offered to improve clinical interpretability, and ensemble saliency maps created through SmoothGrad allow seeing lesion-specific discriminatory areas consistently. Experimental findings indicate that the proposed ensemble is more accurate, robust and minority-class recognizing than individual models, and it has a test accuracy of 89.72%. The framework provides a good, interpretable and computationally efficient solution to automated dermoscopic diagnosis to fit low-resource clinical settings.