Digital Pathology‐Based Comparison of PyRadiomics and HistomicsTK for Nuclei Classification in Melanoma Whole Slide Images

Background: The analysis of histopathological characteristics from biopsy whole slide images (WSI) is a standard procedure in current diagnostic workflows. For instance, malignancies such as melanoma often require the execution of biopsy to be accurately identified. However, diagnosis can be difficult because of variability in clinical scenarios and in microscopic pictures, as well as the lack of biomarkers availability. In this context, the extraction of shape, texture, and intensity-based features from medical images has proven to be a very promising strategy to uncover latent patterns that may be helpful for diagnosis and prediction of several pathologies. Methods: This study proposes radiomics as a powerful tool for extracting nuclei features and enabling nuclei classification of PUMa dataset melanoma WSIs. More specifically, it evaluates the extraction of radiomics features through PyRadiomics, in comparison with the pathomics tool, namely HistomicsTK, in terms of classification performance. To systematically compare these approaches, three supervised classifiers were trained and tested using the same training/testing splits and usual classification metrics: one on radiomics features, one on histomic features, and one on the merged feature set. Results: The results illustrate an improved performance of the radiomics model compared with both the histomic model and the hybrid radiomics and histomics model, suggesting that radiomics can extract valuable phenotypic information from histological images. Conclusions: Radiomics-based feature extraction, as implemented in PyRadiomics, may be a valid and robust alternative to histomics/pathomics descriptors implemented in HistomicsTK in computational pathology pipelines for melanoma analysis.