Coreset Selection for Scalable and Efficient Prostate Cancer Grading in Digital Pathology

Background: Digital pathology generates vast amounts of whole-slide images(WSIs), offering unprecedented opportunities for AI-driven precision medicine.However, the development of robust and generalizable models faces systemic challenges, primarily the scarcity and high cost of high-quality manual annotationsfrom expert clinicians. Efficient selection of representative subsets of datacoresets can reduce annotation effort and computational cost while maintainingmodel performance. This study introduces AIMS (An Informative MeaningfulSubset), a coreset selection strategy and validates it on a use case of prostatecancer (PCa) grading. Methods: This study uses a dataset of 187 H&E-stained whole-slide images(WSIs) from prostatectomy sections, tiled at 20× magnification with 512×512pixel patches. AIMS employs a convolutional autoencoder to learn a compactlatent representation of these tiles, followed by a geometry- and activation-drivensubsampling strategy. The feature space of mean activation vectors is partitioned, and representative coresets (0.5%, 1%, and 2.5% of the full dataset) areselected via cosine similarity. Performance is evaluated using ConvNeXt (trainedfrom scratch) and ResNet18 (with pre-trained UNI embeddings), comparingAIMS-selected subsets to random sampling across datasets with varying characteristics. Model performance is assessed using F1-macro, F1-micro, F1-weighted scores, AUC, and Cohen’s kappa, with additional evaluation of cross-scannergeneralization. Results: Coreset selection via AIMS consistently outperformed random sampling. ConvNeXt trained on AIMS-selected subsets achieved competitive performances using only a fraction of the original data (F1-macro = [0.724-0.782]and AUC = [0.921-0.945]). Cross-scanner evaluation demonstrated robustnessto hardware variability. Similarly, performance were obtained using AIMSselected tiles in the UNI feature space (F1-macro = [0.735-0.742] and AUC =[0.941-0.942]). Conclusion: AIMS-based targeted coreset selection can achieve promising classification performance, representing a viable strategy for substantially reducingannotation costs. This data-efficient approach offers a practical and scalablestrategy for deploying AI in digital pathology, particularly in data-constrainedenvironments, and represents a step toward real-world clinical applicability ofAI-driven PCa diagnosis and grading.