Fig. 1

Study design and workflow. A Schematic diagram of ultrasound images acquisition during diagnosis and treatment of patients with breast cancer who received NAC. Pre-treatment ultrasonography (baseline ultrasonography, denoted as Phase 0), during the first–second cycle of NAC (Phase 1) and posttreatment (Phase 2) ultrasonography were acquired for each patient. B Patients with breast cancer undergoing NAC were divided into the training cohort (YNCH, N = 439, 1317 images) and two external test cohorts (GPPH, N = 212, 636 images, and SPCH, N = 150, 450 images, respectively). The YNCH cohort was used for training an automated tumour segmentation model, and the performance was tested in the GPPH and SPCH cohorts. The ground truth for each image was delineated by experienced radiologists. C Schematic diagram of automated tumour segmentation (U-Net). The middle row of the network output represented the image segmented by the network, and the bottom row represented the ground truth. D Dice similarity coefficient (DICE) of trained automated segmentation model in the training and external test cohorts (large size: ≥ 2cm, small size: < 2cm). The subgroup analysis of DICE was also performed in Phase 0, Phase 1, Phase 2 (E), and different tumour sizes (F). G Images were segmented by the automated segmentation model for feature assessment. Three signatures (P0-Signature, P1-Signature and P2-Signature) were generated and further applied to build the SUAS combined with clinical factors. The performance of SUAS in distinguishing the pathological response (pCR vs. NpCR) was validated in two external test cohorts. Abbreviations: NAC: neoadjuvant chemotherapy; YNCH: Yunnan Cancer Hospital; GPPH: Guangdong Provincial People’s Hospital; SPCH: Shanxi Province Cancer Hospital; DICE: Dice similarity coefficient; SUAS: serial ultrasonography assessment system; pCR: pathological complete response; NpCR: non-pathological complete response