Predicting pathological complete response to breast cancer neoadjuvant therapy using multi-combination machine learning models based on vision transformer features.

The significance of multi-combination machine learning models utilizing vision transformer (VIT) features in forecasting pathological complete response (pCR) after breast cancer neoadjuvant therapy (NAT). A retrospective study was conducted on 124 breast cancer patients who were confirmed by biopsy pathology and underwent surgical resection after NAT to evaluate pCR, and they were divided into a training cohort (n = 87) and a validation cohort (n = 37). Deep learning features were extracted from pre-biopsy ultrasound images based on VIT, ResNet50, and VGG16 convolutional neural network algorithms. Based on the Wilcoxon test, 9, 7, and 7 high-value features were identified from VIT, Resnet50, and VGG16 features, respectively. Using 12 machine learning algorithms, 111, 87, and 82 models were developed utilizing VIT, Resnet50, and VGG16 features. The Stepglm [forward], NaiveBayes, and glmBoost + Ridge ensemble algorithms achieved the highest area under the curve (AUCs). In both the training and validation cohorts, the AUCs of the optimal algorithms were: 0.872 and 0.839 for VIT, 0.837 and 0.804 for Resnet50, and 0.835 and 0.804 for VGG16. The predictive models developed based on VIT features have higher values in evaluating NAT-pCR for breast cancer compared to features quantified by other deep learning algorithms. The application of multi-combination models allows for the selection of the optimal algorithm to achieve higher prediction performance.