Clinical-grade computational pathology using weakly supervised deep learning on whole slide images.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Deep learning for medical imaging

Background:

• Development of pathology decision support systems is limited by the need for large, manually annotated datasets.
• Pixel-wise annotations are time-consuming and expensive, hindering clinical deployment of computational tools.

Purpose of the Study:

• To present a novel multiple instance learning-based deep learning system for pathology.
• To train accurate classification models using only reported diagnoses as labels, bypassing manual annotations.
• To evaluate the system's performance and clinical utility at scale.

Main Methods:

• Developed a multiple instance learning deep learning framework.
• Trained the system using only reported diagnoses as labels on a large dataset of 44,732 whole slide images.
• Evaluated the model on prostate cancer, basal cell carcinoma, and breast cancer metastases datasets without data curation.

Main Results:

• Achieved areas under the curve (AUC) above 0.98 for all tested cancer types.
• Demonstrated the ability to train accurate classification models at an unprecedented scale.
• Clinical application could allow pathologists to exclude 65-75% of slides with 100% sensitivity.

Conclusions:

• The proposed deep learning system effectively trains accurate pathology classification models without manual annotation.
• This approach overcomes a major bottleneck in deploying computational decision support systems in clinical practice.
• The framework lays the foundation for scalable and accurate AI-driven pathology diagnostics.