A Large Dataset to Enhance Skin Cancer Classification With Transformer-Based Deep Neural Networks

The advent of Deep Learning methodologies has revolutionized the field of medical image analysis, particularly in skin lesion diagnosis and classification. This paper proposes an explorative approach utilizing Transformer-based deep neural networks to classify multiclass skin lesion datasets. Initially introduced for natural language processing tasks, Transformers have remarkably succeeded in capturing long-range dependencies in sequential data. However, their application to image data, especially in medical imaging, remains relatively unexplored. Our proposed framework leverages the self-attention mechanism of Transformer models to effectively capture spatial dependencies across image regions without relying on handcrafted features or extensive pre-processing. We present a comprehensive evaluation of several Deep Learning models on skin imaging reference datasets for various types of skin lesions, including melanoma. We objectively evaluate the test performance of the different trained models using a test dataset released in 2023 with ground-truth labels. Our experiments demonstrate that the Transformer-based architecture achieves high performance in lesion classification tasks. The best result was obtained using a Large Dataset, which we modeled by merging smaller datasets, achieving a test accuracy of 86.37%. This dataset can be considered a good solution to improve the generalization capabilities of the Transformer neural network. Our work establishes Transformer-based deep neural networks as a promising framework for skin-lesion classification in medical imaging and potential clinical utility. This research paves the way for further exploration and integration of advanced Deep Learning techniques into medical image analysis, ultimately contributing to a powerful initial analysis tool for clinicians. The code is publicly available at https://github.com/UnluckyMirco/A-Large-Dataset-to-Enhance-Skin-Cancer-Classification-with-Transformer-Based-DNN.