A semi-supervised deep learning and IoMT framework for robust prostate cancer grading under imperfect histopathology data conditions.

Deep learning has shown remarkable promise in histopathological cancer diagnostics; however, its performance heavily depends on the availability of large-scale, high-quality labelled data, which is often an unrealistic assumption in clinical practice. To address the challenges posed by imperfect and limited data, this paper presents a semi-supervised learning (SSL)- driven Internet of Medical Things (IoMT) framework for automated prostate cancer grading using whole-slide histopathology images. The proposed framework integrates deep learning, IoMT data acquisition, and cloud inference to deliver scalable, resource-efficient diagnostics. The framework employs a modified YOLOv11 architecture for region detection and classification, incorporating preprocessing techniques, patch-wise analysis, and mask-based region-of-interest extraction to convert raw biopsy data into structured, clinically relevant Gleason grade predictions (benign, a=Grade 3, 4, or 5). The IoT-enabled imaging devices might support real-time data capture and edge-level preprocessing, while the cloud component facilitates resource efficiency and centralised model inference. This dual-layer framework is designed to support flexible deployment for both on-site diagnostics and remote telepathology workflows, making it highly applicable in under-resourced settings. The framework also incorporates explainability through Grad-CAM to enhance interpretability and clinical trust. Experimental results on the PANDA dataset demonstrate strong diagnostic performance, achieving 81.2% accuracy, an F1-score of 0.68, mAP@0.5 of 0.64, and an average IoU of 0.58. The results suggest that the proposed semi-supervised framework can support prostate cancer grading under imperfect data conditions, while the IoMT architecture provides a basis for future deployment in connected digital pathology workflows.