Machine learning to detect melanoma exploiting nuclei morphology and Spatial organization.

Cutaneous melanoma is one of the most lethal forms of skin cancer, and its incidence is increasing globally. Its diagnosis typically relies on manual histopathological examination, a process that is both complex and time consuming. In this study, we propose an automated diagnostic tool, capable of generating interpretable results to aid clinical decision-making. A total of 146 whole slide images are included in the study, encompassing various lesion types: congenital nevi, dysplastic nevi, melanomas, and melanomas on nevi. The images were first processed using a multi-resolution image processing pipeline with the aim of segmenting nuclei, evaluating their geometrical and morphological features, as well as their spatial organization. To characterize each slide, these features were synthesized into 44 variables, which were then subjected to Linear Discriminant Analysis. Through this procedure, 18 relevant variables were identified demonstrating good performance in melanoma detection, as validated through Monte Carlo Cross-Validation. These variables were also interpreted within the framework of established histopathological diagnostic insights. By refining the analysis to the cellular level, we emulated standard clinical evaluation practices, ensuring that every aspect of the diagnostic process was accessible and verifiable by medical professionals. The proposed tool can offers significant potential to support clinicians in various tasks, such as prioritizing the analysis of critical samples and providing a secondary diagnostic opinion in complex cases.