Ultrasound-based Deep Learning Model For Prediction Of Treatment Response To Neoadjuvant Chemotherapy And Axillary Lymph Node Metastasis In Breast Cancer

Part 1.Ultrasound-based deep learning model for prediction of treatment response to neoadjuvant chemotherapy in breast cancerObjectives: Prediction of pathological complete response(p CR)to neoadjuvant chemotherapy(NAC)prior to treatment in breast cancer is crucial for treatment decision and prognosis evaluation.The study aimed to develop a deep learning model based on ultrasound images to predict pCR to NAC in breast cancer.Methods: A total of 209 patients with pathologically confirmed breast cancer were included in this retrospective study,from November 19,2012 to May 7,2021,including a training set(n=108),an internal validation set(n=28),and an external validation set(n=73).All patients underwent breast ultrasound before biopsy and surgical resection after NAC to assess p CR.Firstly,we transformed 73 low-quality images into high-quality by using super-resolution reconstruction based on U-Net.We proposed a deep learning framework,including a tumor detection network YOLOv5 to detect tumor lesions and a progressive multi-granularity classification network to predict p CR to NAC.Results: For tumor detection task,the mean Average Precision(m AP)of training set,internal validation set and external validation set were 0.975(95%CI: 0.968-0.989),0.892(95%CI:0.875-0.906),0.864(95%CI: 0.851-0.881),respectively.In the prediction task of pCR to NAC,the AUCs of clinical model in the training set,internal validation set and external validation set were 0.876(95%CI: 0.864-0.890),0.693(95%CI: 0.682-0.714),0.620(95%CI: 0.611-0.635),respectively.The AUCs of image model in the training set,internal validation set and external validation set were 0.978(95%CI: 0.962-0.989),0.744(95%CI: 0.728-0.760),0.686(95%CI:0.673-0.698),respectively.The AUCs of combined model were higher than that of the clinical model and image model in the training and validation datasets,were 0.994(95%CI: 0.982-0.998),0.763(95%CI: 0.748-0.779),0.721(95%CI: 0.708-0.736),respectively.Conclusion: Deep learning can capture complex tumor heterogeneity information from ultrasound images,and can be used to noninvasively predict p CR to NAC in breast cancer patients before treatment,providing valuable information for individualized therapy and prognosis assessment.Part 2.Ultrasound-based deep learning model for prediction of axillary lymph node metastasis in breast cancerObjectives: Prediction of axillary lymph node metastasis provides important information for treatment decision and prognosis evaluation in breast cancer.This study aimed to develop a deep learning model based on ultrasound images to predict axillary lymph node metastasis in breast cancer.Methods: A total of 906 patients pathologically confirmed breast cancer were included in this multicenter retrospective study,from March 10,2013 to May 28,2021,including a training set(n=204),an internal validation set(n=52),external validation set 1(n=518),and external validation set 2(n=132).All patients underwent breast ultrasound before biopsy and surgical resection to assess axillary lymph node status.Super-resolution reconstruction based on U-Net was used to transform 518 low-quality images into high-quality images.A tumor detection network YOLOv5 was used to detect tumor lesions and a progressive multi-granularity classification network was used to predict axillary lymph node metastasis.Results: For tumor detection task,the mean Average Precision(m AP)of training set,internal validation set,internal validation set and external validation set were 0.961(95%CI:0.950-0.977),0.880(95%CI: 0.869-0.901),0.849(95%CI: 0.832-0.860),0.835(95%CI:0.824-0.858),respectively.In the prediction task of axillary lymph node metastasis,the AUCs of clinical model in the training set,internal validation set,external validation set 1 and external validation set 2 were 0.731(95%CI: 0.722-0.748),0.648(95%CI: 0.629-0.657),0.652(95%CI:0.628-0.664),0.660(95%CI: 0.643-0.671),respectively.The AUCs of image model in the training set,internal validation set,external validation set 1 and external validation set 2 were0.919(95%CI: 0.907-0.933),0.710(95%CI: 0.718-0.744),0.712(95%CI: 0.698-0.725),0.710(95%CI: 0.697-0.728),respectively.The combined models were superior to clinical model and image model by comparing their predictive performance across all datasets,AUCs were 0.921(95%CI: 0.911-0.936),0.760(95%CI: 0.746-0.772),0.737(95%CI: 0.723-0.751),0.744(95%CI:0.725-0.758),respectively.Conclusion: The prediction model based on deep learning can also be used for noninvasive prediction of the axillary lymph node metastasis in breast cancer,which is helpful to treatment decision and prognosis assessment.