Deep learning methodologies for spitzoid melanocytic tumor characterization

Abstract

The digitization of biopsies into high-resolution whole-slide images has opened the way to artificial intelligence methods in pathology. While histopathological analysis of biopsies remains the gold standard for cancer diagnosis, deep learning holds great potential in reducing pathologist workload and enhancing diagnosis. This can be particularly crucial for tumors with ambiguous morphological features like spitzoid melanocytic lesions, where these methods could greatly improve their clinical interpretation. However, implementing deep learning in digital pathology presents various challenges, encompassing image complexity, limited training data, and the scarcity of annotated samples. Rare cancers can find these limitations amplified, considerably hampering progress in this area. In this work, we aim to tackle the clinical ambiguity surrounding spitzoid melanocytic tumors while addressing the complexities inherent in whole-slide images by leveraging deep learning techniques. To do so, we propose a comprehensive approach combining weakly-supervised learning, federated learning, and active self-training. The developed approaches have been carefully optimized in collaboration with expert pathologists under real-world conditions, to ensure their effectiveness and applicability in clinical settings. By advancing our understanding and utilization of artificial intelligence in digital pathology, we aim to pave the way for improved cancer diagnosis and treatment, particularly in the challenging realm of spitzoid tumors.