Tissue Segmentation: On a holdout dataset, the accuracy for precisely delineating invasive carcinoma tissue structures in images was 92.3%, with a specificity of 99.9% and a sensitivity of 84.7%. In a 3-class scenario - distinguishing between invasive carcinoma, immune, and healthy tissue - the accuracy was 90.0%.
Prediction of pCR: We evaluated the AI using 5-fold cross-validation, achieving a pCR prediction AUC of 0.735±0.017 and an accuracy of 0.68±0.03. The predicted probabilities for responders and non-responders were well-separated, with a p-value of 8.6x10-12.
In our analysis of feature importances, we identified the three most critical feature groups: pathological biomarkers, cell-tissue interactions, and individual cell characteristics. Overall, various AI-based-image-analysis features demonstrated similar levels of importance.
