Preliminary Radiomics-Based Classification of Tumor Nuclei in Melanoma Histopathology

Melanoma is, nowadays, the most aggressive kind of skin tumor. Even if advanced therapies and diagnosis techniques have been developed, metastatic cases still have a poor prognosis. Diagnosis of melanoma is complicated and confounded by the histopathological heterogeneity in samples and the absence of affordable biomarkers. The advent of innovative analytical techniques in medical imaging, such as radiomics and pathomics, poses hope in providing the needed assistance in diagnostic accuracy. In this paper, a preliminary radiomic workflow has been proposed for a subset of 100 annotated Regions of Interest (ROIs) from the PUMA dataset, comprising primary and metastatic melanoma samples. The goal was to differentiate tumor nuclei from non-tumor nuclei using radiomic features extracted from Whole Slide Images (WSIs). Feature extraction was performed using PyRadiomics, and LASSO regression was applied to carry out feature selection and reduce the features count from 102 to 57. The selected features were used to train a Random Forest classifier, whose output was assessed in terms of accuracy, precision, recall, specificity, F1-score, and area under the ROC curve (AUC). The model exhibited 74.56% accuracy, 73% precision, 77% recall, 72% specificity, 75% F1-score and an AUC of 0.82, which indicates a good discriminatory power. The results validated the capability of radiomics in classifying nuclei on melanoma WSIs. Even if limited by using a subset of the PUMA dataset and pre-annotated nuclei, this pipeline is significant in laying the foundation for further integration of radiomics and Artificial Intelligence (AI) in digital pathology.