Data-efficient morphological deep learning for fine-grained Gleason grading based on AI-triaged 3D pathology

Prostate cancer risk assessment relies on 2D histology, which samples only a small portion of tissue and lacks 3D architectural context, often leading to ambiguities. Slide-free 3D pathology using open-top light-sheet microscopy provides volumetric views of intact biopsies, offering richer morphological information without destructive sectioning. Yet the massive scale of 3D datasets makes manual review impractical. Prior work has shown that AI-triage models can identify high-risk 2D sections within 3D pathology datasets for time-efficient pathologist review, but these models are trained on limited labeled datasets, and thus prone to overfitting, especially for fine-grained multiclass tasks (e.g., Gleason grading). To address this, we investigate whether data-efficient, segmentation-guided prototype learning can better capture morphological patterns for improved Gleason grading. To enable fine-grained Gleason grading under limited supervision, we developed SCOPE, a Segmentation-guided CrOss-slice PrototypE learning framework for 3D pathology. Prototype learning helps mitigate overfitting by mapping large numbers of patch features into compact prototype features, each of which represents a distinct tissue morphology. To adapt this paradigm to 3D pathology, we first pretrain on unlabeled 3D volumes to initialize prototypes that capture the broad morphological diversity of the cohort. We then use 3D segmentation masks as structural priors to guide prototype refinement. In addition to the prototype learning, we adopt a 2.5D multiple-instance learning (MIL) strategy that incorporates context from neighboring slices to improve the feature quality for each slice of interest. Aggregated prototype features from each slice are used to train classifiers for grading. To evaluate our approach, we applied SCOPE to a cohort of 59 prostate cancer patients with slice-level annotations (Gleason grades) established by consensus among three pathologists. Using a leave-one-out cross-validation protocol, our framework achieved an AUC of 0.819 for the fine-grained Gleason grading task (i.e., GG1, GG2, GG3, or GG > 3). In comparison, the baseline prototype learning model achieved an AUC of 0.699. We further quantified the contribution of each component in SCOPE, finding that segmentation-guided features contribute the most, followed by the 2.5D MIL formulation and clustering-based pretraining. In summary, SCOPE confirms that data-efficient, segmentation-guided prototype learning can capture fine-grained 3D morphological patterns that improve prostate cancer Gleason grading. By integrating clustering-based pretraining with segmentation-derived structural priors and a 2.5D MIL formulation, SCOPE delivers a high-performance, interpretable, and clinically actionable AI-triage tool for 3D pathology when labeled data is limited.