Development of Fully Automated Deep Learning-based Approach for Prediction of Sentinel Lymph Node Metastasis in Breast Cancer Patients Using Ultrasound Imaging

Taixia Liu^1,2,3Guojia Zhao⁴Wei Wei⁵Qingling Zhang⁵Jing Wu⁵Xuying Chen⁵Dong Liu⁶Xiangming Zhu^1,2,5*

• ¹Cheeloo College of Medicine, Shandong University, Jinan, Shandong, China
• ²Department of Ultrasound, The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, China
• ³Department of Ultrasound, Shanghai Municipal Hospital of Traditional Chinese Medicine, Shanghai, China
• ⁴Department of Ultrasound, Lin Yi People’s Hospital, Linyi, Shandong, China
• ⁵Department of Ultrasound, The First Affiliated Hospital of Wannan Medical College, Wuhu, Anhui, China
• ⁶Department of Information, The First Affiliated Hospital of Wannan Medical College, Wuhu, China

Purpose: This study aimed to develop a novel predicting model based on deep learning (DL) to predict sentinel lymph node (SLN) metastasis in breast cancer (BC) patients using ultrasound (US) imaging.: A retrospective cohort consisting of 692 female BC patients from two hospitals was analyzed. The subjects were categorized into three groups, including a training set (n = 405), an internal validation set (n = 174), and an external testing set (n = 113), respectively. A postfusion model integrating the DeepLabV3, U-Net, and U-Net++ segmentation algorithms, respectively, was utilized to automatically delineate regions of interest (ROIs). Furthermore, three convolutional neural networks (CNNs) -ResNet50, ResNet101and DenseNet121, respectively, were employed to analyze the cropped regions and concurrently construct a predictive model, with ResNet50 identified as the most effective deep learning model. A composite model that incorporates the DL signature (DL Sig) alongside clinical factors was developed by utilizing logistic regression (LR). A database for comparing human and machine performance was created to evaluate the model's effectiveness. A nomogram was ultimately constructed to forecast the occurrence of SLN metastasis. The assessment of model performance involved the utilization of receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA), respectively. Results: The automated DL-based segmentation demonstrated a robust correlation with manual delineation, yielding Dice coefficients of 0.893 and 0.855 in the internal validation and external testing sets, respectively. The ResNet50 model, recognized as the most effective base model, achieved an AUC of 0.765 (95% CI 0.674-0.856) and an accuracy of 74% in the external testing set. The integrated model, which combined the DL Sig with clinical factors , exhibited the most effective performance in forecasting SLN metastasis, achieving an AUC of 0.763 (95% CI 0.671-0.855) and an accuracy of 69% in the testing set. The DCA demonstrated notable clinical utility in the integrated model, surpassing the performance of both senior and junior radiologist clinical translational potential. Conclusions: Our novel predicting model demonstrated advantages over senior and junior radiologists in predicting SLN metastasis, suggesting its potential application in clinical diagnosis thanks to its automatic segmentation and prediction.