Prediction of neoadjuvant therapy response to HER2-positive and triple-negative breast cancer: a multicenter...

In the proof-of-concept study, a deep learning framework based on mammography, multiparametric MRI and clinical features was examined as a tool to predict pathological complete response after neoadjuvant therapy in patients with human epidermal growth factor receptor 2-positive and triple-negative breast cancer.

The retrospective study analyzed 359 breast cancer patients from two institutions. Six unimodal deep learning models (i.e., ADC, DCE-MRI first-phase enhancement, SPAIR T2WI, DWI, CC, MLO) were constructed based on the DenseNet169-CBAM algorithm. These models were subsequently integrated to develop an imaging fusion model. A clinical model was developed using a Multi-Layer Perceptron, with input features that were selected based on univariate and multivariate analyses. A clinical-imaging fusion model was developed by integrating six unimodal deep learning models with selected clinical features. Model performance was evaluated using the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity. The calibration of the predictive models was assessed using calibration curves, and decision curve analysis was performed to evaluate their clinical utility.

Among other unimodal models, ADC deep learning model had excellent performance with AUC values of 0.927 (training), 0.708 (validation), and 0.793 (test). The imaging fusion model achieved AUCs of 0.901 (training), 0.773 (validation), and 0.722 (test). The clinical model achieved AUCs of 0.891 (training), 0.886 (validation), and 0.724 (test). Furthermore, the clinical-imaging fusion model exhibited superior predictive performance, achieving AUCs of 0.992 (training), 0.877 (validation), and 0.768 (test).

The clinical-imaging fusion model based on mammography, multiparametric MRI, and clinical features, exhibited high predictive performance for pCR following NAT in patients with human epidermal growth factor receptor 2-positive and triple-negative breast cancers. The performance of the deep learning model demonstrated its potential to assist in clinical decision-making.