Automatic Segmentation Of Organs-at-risk In Cervical Cancer Brachytherapy Based On U-net Convolutional Neural Network

Objective:An automated outline model based on U-net convolutional neural networks was developed to rapidly and accurately contour the uterus and Organs at Risk(OARs)of patients with cervical cancer,thereby significantly reducing the workload of physicians,shortening the waiting time for patient treatment and ensuring consistency of treatment.Methods:In this study,124 sets of positional computed tomography(CT)images were included.The uterus and OARs(bladder,rectum,sigmoid colon,pelvis,right and left femoral bones)were manually outlined layer by layer according to the International Commission on Radiation Units and Measurements Report No.89(ICRU89).The outline is reviewed and modified by a senior physician.The modified CT images were imported into Accu Learning for model training to obtain an automatic outline model.The model was trained using Dice similarity coefficient(DSC),Hausdorff distance(HD95),Average symmetrical surface distance(ASSD),Relative absolute volume difference(RAVD)geometric metrics to assess their outlining performance.Two senior physicians rated the AI-automated images of 10 patients(0 rejected;1 major revision;2 minor revision;3 no revision required).The AI-assisted scoring time was compared to the manual scoring time of low,medium and high level physicians.The consistency of the patient’s body position and relative position of the applicator,bladder and rectal filling,and dose distribution before and after treatment was compared in the context of shorter treatment times.Correlation analysis was also performed on data with differences before and after treatment.Results:The DSC values of this study model in the seven endangered organs(sigmoid colon,uterus,rectum,bladder,pelvis,left femur,and right femur)were(0.92±0.04,0.87±0.09,0.89±0.07,0.95±0.05,0.99±0.01,0.99±0.01,0.98±0.02),and HD95 values(mm)(7.96±5.61,7.25±4.75,10.81±10.25,6.89±7.29,0.62±0.69,1.25±1.3,2.04±2.82),respectively,and ASSD values(mm)of(1.59±1.32,1.98±1.59,2.07±2.36,1.39±1.77,0.09±0.07,0.17±0.14,0.27±0.29)and RAVD% were(12.31±7.74,16.81±12.89,14.94±17.94,5.33±6.88,1.19±1.5,1.83±1.24,2.77±2.89).The two senior physicians felt that only a small number of the automatically outlined organs in jeopardy required significant modification and that most of them were well suited for clinical use.AIassisted outlining reduced outline time by 42 min,18 min and 10 min for junior,intermediate and senior practitioners respectively,and reduced contouring time by 23 min for radiotherapists on average.Changes in bladder volume,D90(dose at 90%volume),D1cc(maximum dose at 1cm3 volume)and D100(dose at 100% volume)occurred in the same patient at the same treatment plan after a mean interval of45min30s(p < 0.05).There was no significant change in dose in the target area(p >0.05).Correlation analysis showed no significant correlation between bladder D90,sigmoid D1 cc,D100 and bladder volume.High-risk clinical target volume(HR-CTV)volume was significantly correlated with colorectal and bladder D1 cc and D2cc(maximum dose for 2cm3 volume).Conclusion:This study successfully constructed an automatic outlining model for brachytherapy of vital organs in cervical cancer based on U-net convolutional neural network.The clinical evaluation concluded that the outlining results of the model met the needs of clinical application,reduced the outlining time of physicians during brachytherapy,and reduced the waiting time of patients for treatment.The consistency of the bladder,rectum,sigmoid and HR-CTV volumes and doses before and after treatment was evaluated and it was concluded that there was no significant change in dose to the target area before and after treatment with the AI-assisted automated outline model,the source location remained consistent,there was a change in bladder volume and the dose to the organs at risk was reduced,but further support from larger sample sizes and clinical data is needed.In addition,the irradiated dose to organs at risk from brachytherapy for cervical cancer was significantly correlated with the volume of HRCTV.This indicates the importance of individualised efficient and precise segmentation of the target area.