Automatic Delineation Of The Clinical Target Volume And Organs At Risk In Rectal Cancer For Radiation Therapy Using Three-Dimensional Fully Convolutional Neural Networks

Rectal cancer is one of the forms of cancer that kills the highest amount of people every year.As a disease,rectal cancer is highly complex with many factors deciding the outcome of the patient’s survival.The complexity mandates a multi-modality approach in its treatment.Current consensus in the medical community is to attack the cancer with a combination of surgery,chemotherapy,and radiation therapy.The exact combination is a point of controversy and largely depend on the cancer spread.Radiation therapy has nonetheless been shown to significantly reduce local recurrence rate in late stage rectal cancer if administered preoperatively.A major difficulty in radiation therapy planning is the proper delineation of the cancerous tissue for energy beam focus.Clinical target volume(CTV)delineation is riddled with problems such as poor reproducibility and high time costs.The process of target volume delineation is traditionally done by radiation oncologists with years of experience.To alleviate their work burden,automatic segmentation software were developed.Commercially available software for this purpose are currently dominated by algorithms based on one or many atlases used for registration to estimate the target volumes.Emergent deep learning algorithms,like convolutional neural networks(CNNs),that are trained in a supervised fashion show potent results in medical image segmentation tasks.However,their application in radiation therapy planning is scarce.The few networks that have been investigated all utilize two-dimensional kernels.To improve the performance,a three-dimensional network is therefore used in this work.The West China Hospital provided the study with 379 CT scans with accompanying ground truth delineations of relevant region of interests(ROIs)performed by experienced radiation oncologists.In this workflow,65%of the provided CT scans were used for training of the network and the remaining 35%for testing.The 3D CNN used was the V-net and to test its feasibility,it was compared with two other state-of-the-art CNNs using 2D convolutions,the U-net and the DDCNN.The V-net achieved the best overall performance in terms of the mean Dice Similarity Coefficient(DSC),the mean Hausdorff distance(HD)and mean time cost on the testing dataset.The ROIs with corresponding ground truth delineations were the CTV and Organs at risk(OARs)such as the bladder,the small intestine,the colon,and the femoral heads.For the V-net,the range of each performance metric were 0.614-0.928 for the DSC,16.319-3.720 voxels for the HD and 0.7-0.3 seconds in time cost for each ROI.The 3D CNN proposed in this work shows great potential to help clinicians in their treatment of rectal cancer and to improve the patient outcomes throughout the world.