A Convolutional Neural Network Based on Ultrasound Images of Primary Breast Masses: Prediction of Lymph-Node Metastasis in Collaboration With Classification of Benign and Malignant Tumors

Chunxiao Li; Yuanfan Guo; Liqiong Jia; Minghua Yao; Sihui Shao; Jing Chen; Yi Xu; Rong Wu

doi:10.3389/fphys.2022.882648

A Convolutional Neural Network Based on Ultrasound Images of Primary Breast Masses: Prediction of Lymph-Node Metastasis in Collaboration With Classification of Benign and Malignant Tumors

Front Physiol. 2022 Jun 2:13:882648. doi: 10.3389/fphys.2022.882648. eCollection 2022.

Authors

Chunxiao Li¹, Yuanfan Guo², Liqiong Jia³, Minghua Yao¹, Sihui Shao¹, Jing Chen¹, Yi Xu², Rong Wu¹

Affiliations

¹ Department of Ultrasound, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
² Shanghai Key Lab of Digital Media Processing and Transmission, Shanghai Jiao Tong University, Shanghai, China.
³ Department of Ultrasound, Zhongshan Hospital Wusong Branch, Fudan University, Shanghai, China.

Abstract

Purpose: A convolutional neural network (CNN) can perform well in either of two independent tasks [classification and axillary lymph-node metastasis (ALNM) prediction] based on breast ultrasound (US) images. This study is aimed to investigate the feasibility of performing the two tasks simultaneously. Methods: We developed a multi-task CNN model based on a self-built dataset containing 5911 breast US images from 2131 patients. A hierarchical loss (HL) function was designed to relate the two tasks. Sensitivity, specificity, accuracy, precision, F1-score, and analyses of receiver operating characteristic (ROC) curves and heatmaps were calculated. A radiomics model was built by the PyRadiomics package. Results: The sensitivity, specificity and area under the ROC curve (AUC) of our CNN model for classification and ALNM tasks were 83.5%, 71.6%, 0.878 and 76.9%, 78.3%, 0.836, respectively. The inconsistency error of ALNM prediction corrected by HL function decreased from 7.5% to 4.2%. Predictive ability of the CNN model for ALNM burden (≥3 or ≥4) was 77.3%, 62.7%, and 0.752, and 66.6%, 76.8%, and 0.768, respectively, for sensitivity, specificity and AUC. Conclusion: The proposed multi-task CNN model highlights its novelty in simultaneously distinguishing breast lesions and indicating nodal burden through US, which is valuable for "personalized" treatment.

Keywords: breast neoplasms; lymph nodes; multi-task; neural networks; ultrasonography.