Machine learning-based quantitative pathology reveals intratumoral heterogeneity in ductal carcinoma in situ and predicts invasive upstaging.

This study aimed to develop a machine learning framework for objective ductal carcinoma in situ (DCIS) grading and to investigate intratumoral heterogeneity as a predictive biomarker. The grading framework was developed using 267 DCIS resections and validated on two external test sets (n=116 and n=69). A single-cell grading model graded individual nuclei using quantitative measures of nuclear size and pleomorphism. An integrated grading model then combined single-cell with architectural features in a support vector machine classifier. Separately, an invasion prediction model was trained on 288 intermediate- and high-grade DCIS cases and validated on 54 cases to predict upstaging to invasive carcinoma >1 mm. Compared to low-grade nuclei, high-grade nuclei were significantly larger and less uniformly oriented (all P<0.001). The single-cell grading model achieved F1 scores of 0.76, 0.62, and 0.80 for low-, intermediate-, and high-grade nuclei, respectively. In one external test set, the integrated grading model achieved F1 scores of 0.83, 0.71, and 0.72 for low-, intermediate-, and high-grade DCIS, respectively. On 54 intermediate- and high-grade DCIS biopsy cases, combining clinical variables with quantitative heterogeneity metrics improved the prediction of upstaging to invasive carcinoma >1 mm over the clinical-only model, increasing the area under the receiver operating characteristic curve from 0.704 to 0.827. These findings identify quantitatively measured intratumoral heterogeneity as a biologically relevant feature that may explain diagnostic ambiguity in borderline cases and help predict extensive invasive upstaging.