Evidence-Centered Histopathological Grading of Prostate Cancer : A Review of Machine Learning, Hybrid CNN–Transformer Models, Fusion, and Explainable AI

Automated Gleason grading of prostate biopsy histopathology has progressed from handcrafted colour and texture descriptors with conventional classifiers to deep convolutional networks, ensembles, attention and transformer models, hybrid CNN–ViT pipelines, and federated and multi-institutional validation efforts. Recent studies report very high tile-level accuracy, but some persistent problems remain: the grade classes are unequal in size and adjacent grades are hard to separate, evaluation is dominated by accuracy, patient-level, slide-level and external testing is uncommon, and the quality of explanations is rarely measured quantitatively. In this review, we categorize grading methods into handcrafted and classical machine learning, convolutional and transfer-learning systems, ensemble and label-noise-aware models, attention, transformer and hybrid pipelines, and fusion, multimodal and federated approaches. Comparative analysis shows that reported performance is strongly shaped by dataset composition, image resolution, label definition (Gleason pattern versus ISUP grade group), preprocessing and split design, so architectural contributions cannot be ranked from headline figures alone. Evidence-centered grading is therefore not a single network but a combination of complementary local and global representations, grade-aware prediction, calibrated uncertainty, and clinically grounded explainable AI. Priority directions identified in this review include leakage-safe patient-level splitting, slide-level aggregation, multi-centre external validation, calibration reporting, quantitative assessment of explanations, and integration of histology with clinical evidence.