Weakly supervised learning on unannotated H&E-stained slides predicts BRAF mutation in thyroid cancer with high...

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Deep neural networks (DNNs) can predict cancer mutational status from H&E slides, but expert annotations are unreliable for identifying mutation-specific regions.
• This limitation impedes prognostic accuracy and the discovery of new pathobiological insights.

Purpose of the Study:

• To develop and validate a weakly supervised DNN model for predicting BRAF V600E mutational status in thyroid cancer using H&E images.
• To enable precision oncology through inexpensive and timely mutational status prediction without regional annotations.

Main Methods:

• A weakly supervised learning technique trained a DNN on H&E-stained thyroid cancer tissue microarrays (discovery cohort: 85 patients).
• The model's performance was evaluated on an independent external cohort (444 patients).
• A visualization technique was developed to highlight informative regions identified by the DNN.

Main Results:

• The DNN achieved an area under the receiver operating characteristic curve (AUC) of 0.98 (95% CI 0.97-1.00) on the external cohort, significantly exceeding previous DNNs trained with strong supervision.
• The visualization technique provided spatially granular highlighting of informative regions, contributing to explainable artificial intelligence.
• A logistic regression classifier based on pathologist-identified features achieved a lower AUC of 0.78, demonstrating the DNN's superior performance.

Conclusions:

• Weakly supervised learning is effective for training DNNs in scenarios where informative visual patterns are unknown a priori.
• This approach holds significant potential for advancing precision oncology and discovering novel pathobiological knowledge.
• The developed DNN model offers a highly accurate and explainable method for predicting BRAF V600E mutational status in thyroid cancer.