Publication

Abstract : Deep learning methods are transforming the era of medical diagnosis, particularly within oncology, through fully automatic and accurate cancer lesion classification. This study provides a comparativeanalysis of threedeep learning models,the Vision Transformer (ViT), InceptionV3, and Xception, implemented on binary classification of skin cancer, classes named benign and malignant lesions using the open source KaggleMelanoma Skin Cancer datasetconsisting of 10,000 images among which 9600 images are for training the model and 1000 images for evaluation of model.Before using the raw dataset, the images were preprocessed using methods likeresizing, normalization, and data augmentation to enhance generalization. CNN-based models such as InceptionV3 and Xception have their hierarchical feature extraction processes, whereas ViT uses self-attention mechanisms that capture longrange dependencies on the images present in the data, which are essential for medical image analysis. All the models were trained using the Adam optimizer with a learning rate of 1e-4 for 20 epochs. The class imbalance was alleviated using classweightedcross-entropy loss. The models were evaluated using precision,recall, F1-score, and accuracy, where Xception achieved the highest accuracy of 93%, followed by ViT and InceptionV3 at 92%.The comparison shows that ViT is better in feature interpretability as well as robustness, Xception is better in terms of precision and generalization. Xception offers the best architectural efficiency as compared with other models and provides a better classification performance