A LEAKAGE-FREE TWO-PHASE TRANSFER LEARNING ENSEMBLE FOR BINARY MELANOMA CLASSIFICATION USING EFFICIENTNETB3, DENSENET121, INCEPTIONV3, AND VIT-B16

Abstract

Melanoma, a type of skin cancer, is among the fastest-growing and most lethal cancers worldwide. Automation-based early detection of this disease is very crucial for increasing the survival rates of patients. This paper conducts a thorough comparative study of four pre-trained deep learning architectures: EfficientNetB3, DenseNet121, InceptionV3, and Vision Transformer (ViT-B16) with the aid of a weighted ensemble method for binary classification of melanoma on the Kaggle Melanoma Skin Cancer Dataset containing 10,000 dermoscopic images of 5,000 benign and 4,538 malignant skin lesions. Our work features a well-defined two-stage transfer learning methodology that effectively allows for the prevention of data augmentation leakage by performing stratified splitting before augmentation and significantly improves feature adaptation by resorting to progressive fine-tuning and adaptive learning rate scheduling. Individual model accuracies achieved are: EfficientNetB3 (94.90%), DenseNet121 (95.26%), InceptionV3 (95.11%), and ViT-B16 (96.04%). The weighted ensemble that combines the four models achieves 96.77% accuracy, 0.9886 precision, 0.9656 F1-score, and 0.9949 AUC, exceeding the 95.25% ensemble baseline of Sarıateş and Özbay by 1.52 percentage points on the same dataset. This implies that an effectively designed pipeline can continuously increase the accuracy irrespective of the model architecture, and that an ensemble of complementary CNN and Transformer features leads to even better results than single models.