The Gleason grading of prostate cancer is well
known for being subjective in clinical practice, interpathologist
disagreement frequently going above 30% because of persistent
stromal noise and morphological heterogeneity, as well as the
ongoing difficulty in identifying rare high-risk patterns such as
the cribriform subtype. In this study, we introduce the Hybrid
Swin-MIL Network, an automated 5-class Gleason grading (NC,
GG3, GG4, GG5 and GG4-C) framework which has been
assessed on the SICAPv2 benchmark dataset in order to address
these diagnostic problems. We have incorporated into our
pipeline an annotation-guided masking approach that
successfully shuts off the non-target background stroma so that
feature learning is focused solely on the malignant epithelial
structures while still preserving the surrounding benign
morphology. A new architecture combines an intrapatch token
attention pooling hierarchical Swin Transformer with a parallel
multi-scale CNN branch in order to capture the overall
glandular structure while at the same time keeping the fine
cellular texture. Five separate models were trained using a
patient-stratified five-fold cross-validation method and these
were combined by soft voting during inference. We achieved
95.55% accuracy, together with a 95.58% macro F1-score and
95.81% recall for the aggressive GG4-Cribriform subtype
across (12,081) previously unseen test patches using our
ensemble. Indeed, these results highlight the advantages of using
both selective input masking and dual-branch attention, and
cross-validation ensembling provides a robust, reproducible
route to fine-grained prostate cancer grading.
