Monitoring response to neoadjuvant therapy for breast cancer in all treatment phases using an ultrasound deep learning model

Abstract Purpose: The present study investigated whether deep learning models (DLMs) could replace traditional ultrasound measurement models for predicting pathological responses to neoadjuvant chemotherapy (NAC) for breast cancer. Methods: Data from 57 patients (443 ultrasound images) who underwent NAC followed by surgery were analyzed. A DLM was developed for accurate breast tumor ultrasound image segmentation. The predictive abilities of the DLM, manual segmentation model (MSM), and two traditional measurement models (longest axis model [LAM] and dual-axis model [DAM]) for pathological complete response (pCR) were compared using tumor size ratios and receiver operating characteristic curves. Results: The average intersection over the union value of the DLM was 0.8087. MSM showed the best performance with an area under the curve (AUC) of 0.840; DLM performance was slightly weaker with an AUC of 0.756. The AUCs of the two traditional models were 0.778 for LAM and 0.796 for DAM. There was no significant difference in AUC values of the predictive ability of the four models. Moreover, no significant difference in AUC values of ultrasound prediction was noted between each NAC cycle (p<0.05). Conclusion: Patients in the pCR group had a significantly better response than those in the non-pCR group, and ultrasonography was predictive of pCR in the early stages of NAC. DLMs can replace traditional measurements for predicting pCR.