Early Prediction of Pathological Complete Response to Neoadjuvant Chemotherapy in Breast Cancer using a Longitudinal US-based Stack-model.

RATIONALE AND OBJECTIVES: To evaluate the diagnostic performance of a longitudinal ultrasound (US)-based stack-model for early prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in breast cancer, as well as its practicality in assisting radiologists with diagnostic ability. MATERIALS AND METHODS: A total of 974 patients who underwent NAC were retrospectively included between January 2017 and March 2022 from three different institutions. For all patients, US imaging was performed before NAC and after two cycles of NAC. The patients from the first hospital were used as the training dataset (n=653) and patients from hospital 2 and 3 were used as the test dataset (n=321)to test five deep learning (DL) models based on different feature sets. The optimal model was selected according to the area under the receiver operating characteristic curve (AUC). A model combining US imaging features with clinical factors was also investigated. Furthermore, the applicability of this model to provide clinical assistance was examined by radiologists with varying degrees of seniority. RESULTS: The Swin Transformer model based on the stacked-feature set achieved the highest AUC. Upon incorporating clinical factors, the combined model demonstrated superior performance in predicting pCR, achieving AUC of 0.935. Diagnostic performance in the early prediction of pCR improved for radiologists across all experience levels when assisted by the combined model. CONCLUSION: The longitudinal US-based model enables early prediction of pCR. Additionally, the model provided positive diagnostic assistance to radiologists with different experience levels. CRITICAL RELEVANCE STATEMENT: The longitudinal US-based model enable non-invasive early prediction of response to neoadjuvant chemotherapy in breast cancer while enhancing diagnostic performance across radiologists with varying experience levels.