PathoCoder: Rethinking the Flaws of Patch-Based Learning for Multi-Class Classification in Computational Pathology

Pathology-based decision support systems in clinical settings have faced impediments from data preparation beforehand, large-scale manual annotations, and poor domain generalization. We report a unified hybrid framework with only raw, slide-level label images. The method, which we termed PathoCoder, comprises core feature extractors, a feature combiner/reduction, and a supervised classifier. It is trained (through 5-fold cross-validation) on 2452 SurePath cervical liquid-based whole-slide captures, provided from Mendeley repository. Tests resulted in 98.37%, 98.37%, 98.41%, and 98.37% in accuracy, precision, recall, and F1, respectively. Extensive experiments validate the proposed scheme and versatility enough to accommodate epithelial ovarian tumor histotypes. Our method paves the way for more accelerated advancements in pathology AI by reducing patch/pixel-based annotation and good tissue quality dependency. Its applicability spans diverse classification tasks with varying tissue content and holds potential for real-world implementation.