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

Area of Science:

• Digital Pathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Pathology-based decision support systems face challenges with data preparation, manual annotations, and domain generalization.
• Existing AI models often require extensive, time-consuming, and resource-intensive data preprocessing and annotation.

Purpose of the Study:

• To introduce PathoCoder, a unified hybrid framework for pathology image analysis using only raw, slide-level label images.
• To overcome the limitations of traditional AI approaches in digital pathology by minimizing data preparation and annotation requirements.

Main Methods:

• Developed PathoCoder, a framework integrating core feature extractors, feature combination/reduction, and a supervised classifier.
• Trained and validated the model using 5-fold cross-validation on 2452 SurePath cervical liquid-based whole-slide images from the Mendeley repository.

Main Results:

• Achieved high performance metrics: 98.37% accuracy, 98.37% precision, 98.41% recall, and 98.37% F1-score.
• Demonstrated versatility and validated through extensive experiments, showing applicability to epithelial ovarian tumor histotypes.

Conclusions:

• PathoCoder significantly reduces the dependency on patch/pixel-based annotation and high-quality tissue, enabling faster AI advancements in pathology.
• The framework's applicability to diverse classification tasks and varying tissue content suggests strong potential for real-world clinical implementation.