Pathologist-like explainable AI for interpretable Gleason grading in prostate cancer

GleasonXAI employs a U-Net architecture trained on 1,015 tissue microarray core images. These images were annotated with detailed, pathologist-defined pattern descriptions by 54 international pathologists. A key innovation is the use of 'soft labels' to incorporate intrinsic uncertainty and interobserver variability into the training data, rather than relying on a single 'hard label' majority vote. This approach allows the model to learn from the nuances of expert judgment and provide detailed, interpretable segmentations.

The GleasonXAI model achieved a Dice score of 0.713 ± 0.003 for Gleason pattern segmentation, outperforming direct Gleason pattern segmentation (0.691 ± 0.010). The model's interpretability is derived from its direct prediction of histological features, which are then mapped to Gleason patterns. While agreement among pathologists varied, especially for rarer or more detailed sub-explanations, the soft-label approach allowed the model to robustly handle this variability.

GleasonXAI offers significant advancements in interpretable AI for medical diagnosis. By providing pathologist-like explanations, it enhances trust and facilitates clinical adoption, especially in high-variability tasks like Gleason grading. Future work will focus on expanding data collection for rarer patterns and exploring advanced techniques for handling predictive distributions in segmentation tasks.