A deep learning framework for histopathological classification of canine testicular tumours

Testicular neoplasms are common in dogs, and their histopathological classification may be challenging in selected cases, particularly when morphological patterns overlap, with potential implications for diagnostic consistency and case management. In this study, we propose an artificial intelligence (AI)-based computational pathology approach for the automatic classification of the three main canine testicular neoplasms (Seminoma, Leydig cell tumour, and Sertoli cell tumour) from digitised haematoxylin and eosin-stained Whole-Slide Images (WSIs). To this end, 200 histological WSIs from 200 male dogs with histologically diagnosed testicular tumours were collected from the archives of the DIPSA laboratory, University of Naples Federico II. We fine-tuned the EfficientNet-B4 convolutional neural network (CNN) on tiles from WSIs and evaluated performance using 5-fold cross-validation with WSI-level partitioning. In addition, we implemented an attention-based multiple instance learning (ABMIL) framework as a complementary weakly supervised approach for WSI-level classification. The fine-tuned EfficientNet-B4 model provided the strongest tile-level discrimination, while Logistic Regression with mean probability aggregation achieved the highest WSI-level performance among the evaluated WSI-level approaches (accuracy 0.84 ± 0.07). Furthermore, the ABMIL framework showed competitive WSI-level performance, reaching an accuracy of 0.80 ± 0.04. Finally, Grad-CAM heatmaps revealed that the fine-tuned CNN focused on histologically relevant tumour regions. Overall, our findings support the feasibility of AI-assisted WSI-level histopathological classification of canine testicular tumours and provide a proof-of-concept for the potential development of decision-support tools for veterinary pathologists.