Developing and validating deep learning-based heterogeneous model to improve diagnostic performance of ultrasound elastography for axillary lymph node metastasis of early breast cancer.

Abstract

e12583 Background: To improve the performance of ultrasound (US) for diagnosing metastatic axillary lymph node (ALN), machine learning was used to reveal the inherently medical hints from ultrasonic images and assist pre-treatment evaluation of ALN for patients with early breast cancer. Methods: A total of 214 eligible patients with 220 breast lesions, from whom 220 target ALNs of ipsilateral axillae underwent ultrasound elastography (UE), were prospectively recruited. Based on feature extraction and fusion of B-mode and shear wave elastography (SWE) images of 140 target ALNs using radiomics and deep learning, with reference to the axillary pathological evaluation from training cohort, a proposed deep learning-based heterogeneous model (DLHM) was established and then validated by a collection of B-mode and SWE images of 80 target ALNs from testing cohort. Performance was compared between UE based on radiological criteria and DLHM in terms of areas under the receiver operating characteristics curve (AUC), sensitivity, specificity, accuracy, negative predictive value, and positive predictive value for diagnosing ALN metastasis. Results: DLHM achieved an excellent performance for both training and validation cohorts. In the prospectively testing cohort, DLHM demonstrated the best diagnostic performance with AUC of 0.911(95% confidence interval [CI]: 0.826, 0.963) in identifying metastatic ALN, which significantly outperformed UE in terms of AUC (0.707, 95% CI: 0.595, 0.804, P<0.001). Conclusions: DLHM provides an effective, accurate and non-invasive preoperative method for assisting the diagnosis of ALN metastasis in patients with early breast cancer.[Table: see text]