Abstract 60P
Background
Clinical trials are exploring the possibility of active surveillance for low-risk ductal carcinoma in situ (DCIS) and atypical ductal hyperplasia (ADH). However, interobserver variability in grading DCIS exists. We aim to train and evaluate a deep learning model (Deep DCIS) on breast biopsy with ADH/DCIS to improve the risk stratification of these precursor lesions.
Methods
Patients diagnosed with ADH or DCIS by breast biopsy and who received wide excision within six months were included (n=592 and 112 for primary and independent validation cohorts). Two pathologists independently evaluated nuclear grade, comedonecrosis, and focus suspicious for microinvasion and annotated patch-level labels by consensus. Estrogen receptor immunostain and apocrine features were also assessed. Clinical features (e.g., biopsy method, lesion size, BIRADS classification of ultrasound and mammogram) were collected. The model adopted Inception-V3 with a train-valid-test split and Adam optimization. The outputs of Deep DCIS comprised five parameters: total patches, lesion extent, Deep Grade, Deep Necrosis, and Deep Stroma. These parameters were combined with clinical information and hormone status to predict upstaging to invasive breast carcinoma.
Results
Deep Grade strongly correlated with the pathologists' grading on both slide- and patient-level labels. All five parameters from Deep DCIS were associated with upstaging to invasive carcinoma. Deep DCIS successfully predicted upstaging with an area under the curve (AUC) of 0.81 and an accuracy (acc) of 0.75, outperforming pathologists' evaluation (AUC 0.71 and 0.69; acc 0.67 and 0.65 from two pathologists). The prediction improved after combining Deep DCIS with clinical factors and hormone status (AUC 0.87; acc 0.79). This combined model retained its predictive power in the subgroup of ADH and low-/intermediate-grade DCIS (AUC 0.81; acc 0.76). The independent validation cohort confirmed the model's performance (AUC 0.82; acc 0.72).
Conclusions
The deep learning model refines histological evaluation of ADH and DCIS on breast biopsy and improves prediction of upstaging to invasive carcinoma on wide excision, which may help guide treatment planning in future clinical studies.
Legal entity responsible for the study
Chung-yen Huang.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.