60P Deep learning-based digital pathology for risk stratification of atypical ductal hyperplasia/ductal carcinoma in situ

Clinical trials are exploring the possibility of active surveillance for low-risk ductal carcinoma in situ (DCIS) and atypical hyperplasia (ADH). However, interobserver variability grading DCIS exists. We aim to train evaluate a deep learning model (Deep DCIS) on breast biopsy with ADH/DCIS improve risk stratification these precursor lesions. Patients diagnosed ADH or by who received wide excision within six months were included (n=592 112 primary independent validation cohorts). Two pathologists independently evaluated nuclear grade, comedonecrosis, focus suspicious microinvasion annotated patch-level labels consensus. Estrogen receptor immunostain apocrine features also assessed. (e.g., method, lesion size, BIRADS classification ultrasound mammogram) collected. The adopted Inception-V3 train-valid-test split Adam optimization. outputs Deep comprised five parameters: total patches, extent, Grade, Necrosis, Stroma. These parameters combined clinical information hormone status predict upstaging invasive carcinoma. Grade strongly correlated pathologists' both slide- patient-level labels. All from associated successfully predicted an area under curve (AUC) 0.81 accuracy (acc) 0.75, outperforming evaluation (AUC 0.71 0.69; acc 0.67 0.65 two pathologists). prediction improved after combining factors 0.87; 0.79). This retained its predictive power subgroup low-/intermediate-grade 0.81; 0.76). cohort confirmed model's performance 0.82; 0.72). refines histological improves excision, which may help guide treatment planning future studies.