Multi-center study: ultrasound-based deep learning features for predicting Ki-67 expression in breast cancer.

Applying deep learning algorithms to mine ultrasound features of breast cancer and construct a machine learning model that accurately predicts Ki-67 expression level. This multi-center retrospective study analyzed clinical and ultrasound data from 929 breast cancer patients. We integrated deep features from the tumor and peritumoral areas to build a fusion model for predicting Ki-67 expression. The model underwent performance validation on both internal and external test datasets. Its accuracy as well as clinical usefulness were evaluated by diverse statistical metrics. In the ultrasound depth feature model for the tumor area, the Support Vector Machine (SVM) algorithm achieved the highest performance, with an accuracy of 0.782, ROAUC of 0.771 (95% CI 0.704-0.838), sensitivity of 0.905, specificity of 0.543, and F1 score of 0.846. In the depth feature model for the peritumoral area, the Light Gradient Boosting Machine (LightGBM) algorithm demonstrated superior performance, achieving an accuracy of 0.728, ROAUC of 0.623 (95% CI 0.545-0.702), sensitivity of 0.892, specificity of 0.407, and F1 score of 0.813. The SVM algorithm exhibited superior performance in both internal and external test sets when validated the fusion model integrating depth features from tumor and peritumoral area. Internal test set validation in clinical application indicated significantly lower disease-free survival in the high Ki-67 expression group compared to the low expression group (P = 0.005). Through comprehensive analysis of breast cancer ultrasound images and the application of machine learning techniques, we developed a highly accurate model for predicting Ki-67 expression levels.