Automatic Segmentation Of High-risk Clinical Target Volume Of Cervical Cancer And Organs At Risk For Brachytherapy With A Convolutional Neural Network

AimThis study aimed to designed an auto-segmentation model based on convolutional neural network(CNN),which could achieve delineation of high-risk clinical target volume(HR-CTV)of cervical cancer and pelvic organs at risk(OARs)for brachytherapy.This model would improve the consistency of segmentation and clinical work efficiency,and shorten the patient’s waiting time for treatment.Materials and MethodsThe computed tomography(CT)scans of 68 patients with locally advanced cervical cancer who underwent image-guided brachytherapy(IGBT)were collected.The HR-CTV and OARs including bladder,rectum,sigmoid and intestine were modified manually by a same clinician.Then the results were reviewed and revised by two senior oncologists.A 2.5D CNN model based on 2D U-Net was constructed to achieve auto-segmentation.The Dice similarity coefficient(DSC)and 95th Hausdorff distance(95HD)were used to evaluated its segmentation performance quantificationally.The original contours of the 10 cases in the test set were selected as the ground truth(GT),and then compared with the auto-segmentation contours though a blinded evaluation about the clinical use of the contours by two oncologists.The difference in clinical scoring between the two groups and the two oncologists were compared.In addition,we modified the weight distribution of DSC according to the inverse-square law of BT,and then tested it on the automatic segmentation results of 16 patients with cervical cancer who underwent post-surgery brachytherapy.The values of DSC and weight-DSC(wDSC)were compared,and explored their correlation with D2cc.ResultsThe mean DSC of this auto-segmentation model were 0.81,0.92,0.84,0.65 and 0.83 for the HR-CTV,bladder,rectum,sigmoid and intestine,respectively.The corresponding 95HD were 5.07mm,4.56mm,4.58mm,28.31mm,and 18.77mm,respectively.The evaluation results of the two clinicians showed that most of the bladder,rectum and small intestine contoured by the model meet the clinical application standards,while 70%of the sigmoid contours needed to be greatly modified.98.7%of the HR-CTV contoured by the model could be accepted by the two clinicians,although the mean score of AI group is significantly lower than that of the GT group.The mean scores of HR-CTV scores for AI and GT group of clinician A were significantly lower than those of clinician B.With the assist of AI,the time for manual segmentation was shortened from 28min to 15min.In addition,the mean DSC for the 16 patients was 0.93 for the bladder and 0.87 for the rectum,while the mean wDSC were 0.52 for the bladder and 0.40 for the rectum.The absolute value of the D2cc difference(|△D2cc|)between the GT group and AI group was obtained.The mean |△D2cc| of bladder was 39.94cGy,and the mean |△D2cc| was 42.06cGy.Correlation analysis showed that bladder wDSC and |△D2cc| were significantly correlated,and the correlation coefficient was-0.572(P<0.05).ConclusionThe HR-CTV and OARs of cervical cancer generated by the proposed model could meet the clinical requirements,while a larger training data was still needed to improve the delineation results of the rectum and sigmoid.This model could help improve segmentation efficiency and standardize the clinical work.This study also indicated that there were different delineation preferences and different understandings on the standard of HR-CTV.In addition,the better correlation between wDSC and D2cc showed that it might be more suitable to assess the segmentation performance in BT.