Deep learning for real-time detection of breast cancer presenting pathological nipple discharge by ductoscopy

ObjectiveAs a common breast cancer-related complaint, pathological nipple discharge (PND) detected by ductoscopy is often missed diagnosed. Deep learning techniques have enabled great advances in clinical imaging but are rarely applied in breast cancer with PND. This study aimed to design and validate an Intelligent Ductoscopy for Breast Cancer Diagnostic System (IDBCS) for breast cancer diagnosis by analyzing real-time imaging data acquired by ductoscopy. Materials and methodsThe present multicenter, case-control trial was carried out in 6 hospitals in China. Images for consecutive patients, aged >= 18 years, with no previous ductoscopy, were obtained from the involved hospitals. All individuals with PND confirmed from breast lesions by ductoscopy were eligible. Images from Beijing Chao-Yang Hospital were randomly assigned (8:2) to the training (IDBCS development) and internal validation (performance evaluation of the IDBCS) datasets. Diagnostic performance was further assessed with internal and prospective validation datasets from Beijing Chao-Yang Hospital; further external validation was carried out with datasets from 5 primary care hospitals. Diagnostic accuracies, sensitivities, specificities, and positive and negative predictive values for IDBCS and endoscopists (expert, competent, or trainee) in the detection of malignant lesions were obtained by the Clopper-Pearson method. ResultsTotally 11305 ductoscopy images in 1072 patients were utilized for developing and testing the IDBCS. Area under the curves (AUCs) in breast cancer detection were 0 center dot 975 (95%CI 0 center dot 899-0 center dot 998) and 0 center dot 954 (95%CI 0 center dot 925-0 center dot 975) in the internal validation and prospective datasets, respectively, and ranged between 0 center dot 922 (95%CI 0 center dot 866-0 center dot 960) and 0 center dot 965 (95%CI 0 center dot 892-0 center dot 994) in the 5 external validation datasets. The IDBCS had superior diagnostic accuracy compared with expert (0.912 [95%CI 0.839-0.959] vs 0.726 [0.672-0.775]; p<0.001), competent (0.699 [95%CI 0.645-0.750], p<0.001), and trainee (0.703 [95%CI 0.648-0.753], p<0.001) endoscopists. ConclusionsIDBCS outperforms clinical oncologists, achieving high accuracy in diagnosing breast cancer with PND. The novel system could help endoscopists improve their diagnostic efficacy in breast cancer diagnosis.