Automatic Segmentation Of Clinical Target Volume Of Breast Cancer And Organs At Risk For Radiotherapy With A Convolutional Neural Network

ObjectiveThis study aimed to construct a model based on convolutional neural network,which can fulfill automatic segmentation of clinical target volume(CTV)of breast cancer and organs at risk(OARs)for radiotherapy.We hope this model can shorten the contouring time and provide accurate segmentation results.Materials and MethodsWe collected computed tomography(CT)scans of 160 patients who underwent breast conservative surgery and 110 patients who underwent modified radical mastectomy.They all need to accept adjuvant radiotherapy after operation.CTV and four organs including contralateral breast,heart,lungs and spinal cord were contoured manually.The annotated contours were confirmed by two experienced oncologists.A multi-class segmentation model based on 2D U-Net was designed to fulfill segmentation tasks efficiently.The Dice similarity coefficient(DSC)and 95th Hausdorff distance(95HD)were used as quantitative evaluation metrics to evaluate performance of our model.The CT data of 10 cases of breast conservative surgery and 10 cases of modified radical mastectomy were randomly selected as the ground truth(GT)mask,the corresponding delineation results were generated by artificial intelligence(AI).Then the blinded clinical evaluation was performed by two clinicians in our center.Score differences between these two groups(GT versus AI)and the evaluation consistency were compared.ResultsThe mean DSC of the proposed model are 0.94,0.90,0.95,0.94,0.96,0.96 and 0.93 for the breast CTV,chest wall CTV,contralateral breast,heart,right lung,left lung and spinal cord.The evaluation results of two doctors showed that all OARs contoured by the model can be applied in clinical practice.99.4%of breast CTV and 99.3%of chest wall CTV can be accepted by doctor A,while doctor B was 99.4%and 98.9%.The score differences between GT group and AI group for OARs and CTV showed no statistically significance.The mean scores of OARs in AI group given by doctor B were all higher than doctor A,but the situation was opposite for the CTV.For the heart,right lung and CTV,the differences between two doctors were statistically significant.The consistency coefficient Kappa was 0.266.It took 6.15 s,3.88 s,3.45 s to delineate OARs,breast CTV and chest wall CTV separately by the model,and can help to lessen oncologists’ efforts.ConclusionOur proposed model can generate CTV of breast cancer and OARs which meet the clinical requirements and have same quality as human-generated contours.This model can help reduce clinical workload and make clinical work more standardized.The study also indicated that different observers have different preference and have different understanding about the contouring principles.At the same time,the research also provides a new idea for construction and evaluation of medical image segmentation model in the future.