Predicting pathological complete response to breast cancer neoadjuvant therapy using multi-combination machine...

Area of Science:

• Oncology
• Medical Imaging
• Machine Learning

Background:

• Accurate prediction of pathological complete response (pCR) after neoadjuvant therapy (NAT) is crucial for breast cancer management.
• Deep learning models, particularly those using vision transformer (VIT) features, offer potential for improved prediction accuracy.

Purpose of the Study:

• To evaluate the efficacy of multi-combination machine learning models utilizing VIT features for forecasting pCR in breast cancer patients undergoing NAT.
• To compare the predictive performance of models based on VIT features against those using ResNet50 and VGG16 features.

Main Methods:

• A retrospective study included 124 breast cancer patients. Deep learning features were extracted from pre-biopsy ultrasound images using VIT, ResNet50, and VGG16.
• High-value features were identified using the Wilcoxon test. Multiple machine learning algorithms were employed to develop predictive models.
• Model performance was assessed using the area under the curve (AUC) in training and validation cohorts.

Main Results:

• Multi-combination models using VIT features achieved the highest AUCs (0.872 training, 0.839 validation).
• Optimal algorithms included Stepglm [forward], NaiveBayes, and glmBoost + Ridge ensembles.
• VIT-based models demonstrated superior performance compared to ResNet50 and VGG16 in evaluating NAT-pCR.

Conclusions:

• Machine learning models incorporating VIT features are effective for predicting NAT-pCR in breast cancer.
• Multi-combination approaches enable the selection of optimal algorithms for enhanced predictive accuracy.
• VIT features offer a valuable tool for improving breast cancer treatment response assessment.