Transformer-Based Skin Carcinoma Classification using Histopathology Images via Incremental Learning

This work introduces an innovative transformer-based technique for the classification of non-melanoma skin cancer by the analysis of histopathology images in an incremental learning setting. AI, particularly with the emergence of deep learning and CNN models, have been very impactful in the past decade for medical image analysis. Consequently, research has also emerged for the diagnosis of skin cancer but mostly based on dermoscopic or clinical images. Skin cancer is, however, diagnosed by pathologists after a critical evaluation of histopathology slides extracted through biopsies. The availability of large-scale labelled datasets of histopathology images for the training of models is a challenge which can be resolved by incremental learning and the models can be incrementally trained with the availability of new data or on the discovery of a new disease. Moreover, transformer-based image analysis approaches, with their ability to capture global context, have recently demonstrated superior performance. Considering histopathology slides as the gold standard for skin cancer diagnosis this research, therefore, employs a novel transformer-based classification approach in an incremental learning setting to classify multi-class skin cancer. The method has been validated using a recently published histopathology dataset with superior classification results in comparison with known models.