Prediction model of axillary lymph node metastasis of breast cancer（≤2.5 cm） based on deep learning ultrasound features
10.3760/cma.j.cn131148-20250428-00240
- VernacularTitle:基于深度学习超声特征构建乳腺癌（≤2.5 cm）腋窝淋巴结转移的预测模型
- Author:
Yuyang GAN
1
;
Dongming WEI
;
Ruilong YAN
;
Haiman SONG
;
Jia LI
;
Ziyi YIN
;
Tao CHEN
;
Tengfei YU
Author Information
1. 首都医科大学附属北京天坛医院超声科，北京 100070
- Publication Type:Journal Article
- Keywords:
Ultrasonography;
Breast cancer;
Axillary lymph node metastasis;
Deep learning;
Prediction model
- From:
Chinese Journal of Ultrasonography
2025;34(9):751-758
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To establish a model based on the characteristics of breast cancer ultrasound images through deep learning methods to predict the risk of axillary lymph node metastasis（ALNM）in patients with breast cancer（maximum diameter ≤2.5 cm）before surgery.Methods:A total of 419 patients（3 433 breast tumor ultrasound images）with breast cancer（maximum diameter ≤2.5 cm）who underwent axillary lymph node dissection at Beijing Tiantan Hospital，Capital Medical University from January 2019 to December 2024 were retrospectively included. According to the pathological results of axillary lymph nodes，they were divided into 220 cases in the ALNM occurrence group（positive group）and 199 cases in the non-ALNM occurrence group（negative group）. The breast cancer ultrasound images of the two groups of cases were randomly classified into the training set（2 404 images），the validation set（687 images）and the test set（342 images）according to a ratio of 7∶2∶1. YOLOv8 was used as the basic model of You Only Look Once（YOLO）and optimized. The optimized model was applied to locate and capture the potential ultrasound features of breast cancer cases in the training set. A prediction model was constructed based on the captured ultrasound features. The model was adjusted and optimized through the validation set，and then matched with the case images in the test set. The confusion classification matrix graph and the curve graph for measuring the model performance were used to evaluate the model prediction performance and interpret the model，and the efficacy of this model in identifying breast cancer patients at risk of ALNM was analyzed.Results:There were statistically significant differences between the positive and negative groups in terms of the pathological maximum diameter of breast tumors，pathological T staging，the differentiation degree，the presence of distant metastasis，the maximum diameter measured by ultrasound，the quadrant of breast tumor occurrence，the Breast Imaging - Reporting and Data System（BI-RADS）classification of breast tumors，and the presence of abnormal ultrasound features of lymph node（all P<0.05）. The established deep learning model could automatically perform bounding box localization for the breast cancer of patients.The breast tumors in the positive group had potential ultrasound features that could be captured by the model compared with those in the negative group. The mean average precision（mAP）50 was 0.883，mAP 50-95 was 0.636，PR-AUC was 0.884 5，strict PR-AUC was 0.636 4，the sensitivity was 90.5%，and the specificity was 91.2%，and it had a good predictive efficacy. Conclusions:This prediction model based on the ultrasound characteristics of breast cancer through deep learning can effectively predict breast cancer（maximum diameter ≤ 2.5 cm）with the risk of ALNM，providing an effective basis for the clinical management of axillary lymph nodes in breast cancer patients.
