Background <p>Recent advances in histology scanning technology and Artificial Intelligence (AI) offer great opportunities to support cancer diagnosis. The inability to interpret the extracted features and model predictions is one of the major issues limiting the acceptance of AI models in clinical practice, and a clear representation of the relevance of the extracted features and model predictions is lacking. Focusing on the problem of prostate cancer (PCa) diagnosis and grading, this study aims to detect which are the most discriminant features for distinguishing malignant from non-malignant tissue and Gleason patterns, leaving the evaluation of models’ classification performances as a secondary goal.</p> Methods <p>Utilizing a dataset of 187 annotated H&amp;E-stained whole-slide images, the study explores three magnification levels, extracting 1971 features per tile. Two machine-learning classification tasks are conducted (Malignant vs. Non-malignant and High-grade vs. Low-grade) for each magnification. SHapley Additive exPlanations method was used to investigate models’ interpretability, estimating the importance of pathomic features and their impact on prediction models.</p> Results <p>Wavelet features were consistently prominent in “High-grade vs Low-grade” classification task, together with local binary pattern descriptors. Some histogram features appeared as key features for diagnostic classification tasks. The identified key discriminant features were classified for their specificity with respect to WSI magnification. Very high AUC values were reached for PCa diagnosis task (0.97 &lt; AUC &lt; 0.99), while task involving Low- versus High-grade classification exhibit lower AUC values (maximum AUC: 0.72–0.73).</p> Conclusions <p>This work provides new insights for the explanation of hand-crafted pathomic-based classification models for PCa diagnosis and grading.</p> Clinical trial number <p>Not applicable.</p>