Deep Learning Radiomics Based on Ultrasound Images for Predicting Postoperative Brain Metastasis Risk In Breast Cancer Patients: A Multicenter Study
Abstract
Purpose
The purpose of this study is to utilize clinical pathological information and preoperative ultrasound images of breast cancer to predict the risk of brain metastasis by combining radiomics and deep learning.
Methods
A retrospective design was adopted, and patients with primary breast cancer from two hospitals were included. Preoperative ultrasound images and clinical pathological information were collected, ROI regions were outlined, and radiomics features were extracted and screened to construct a radiomics model. After preprocessing the images, a deep learning model was established, and the optimal deep learning model was obtained. The clinical pathological information, radiomics scores, and the predicted scores from the optimal deep learning model were combined and used for machine learning to establish a fusion model for evaluating the risk of brain metastasis in breast cancer.
Results
This study collected a total of 361 patients diagnosed with primary breast cancer and undergoing surgery in two hospitals from 2012 to 2021, of which 91 patients developed brain metastasis (positive rate of 25.21%). In the radiomics model, the AUC result of the XGBoost prediction model combined with clinical pathological information reached 0.756 in the external validation cohort; in deep learning, the best deep learning model was the ResNet 34 Into ViT model within the CNN-ViT combined architecture, achieving a final AUC result of 0.713 in the external validation cohort; the fusion model, which combines clinical pathological information, radiomics score (Rad Score), and the best deep learning model score (DL Score), achieved the best prediction results, with an AUC of 0.859 in the external validation cohort.
Conclusion
The integration of radiomics, deep learning, and clinical information can effectively enhance the accuracy of predicting the risk of brain metastasis after primary breast cancer surgery, providing auxiliary support for clinical decision-making.