Automatic Gas Detection for Prostate Cancer Patient on Image-Guided Radiation Therapy Using a Deep Convolutional Neural Network

Abstract

Detection of the intestinal/rectal gas is very important task for prostate cancer patient on image guided radiation therapy (IGRT). An automatic intestinal/rectal detection system for prostate cancer patient can reduce the workload of therapists and assist them in carrying out safe and reliable radiotherapy treatment. We proposed a deep convolutional neural network (DCNN) to detect the intestinal/rectal gas in pelvis region. KV X-ray 300 images were acquired from 30 prostate cancer patients with simultaneously acquired on both of the irradiation port at 0 and 90 degrees. 30 images were randomly chosen as tested set and the remaining 270 images as training set. In this study, we used the kV image of 0 degree. Image matrix size was 1024 × 768 pixels. The intestinal/rectal gas was manually delineated on kV X-ray images and segmented. The training subjects were augmented by applying artificial shifts and feed to a convolutional neural network. All images were scaled to a size of 256 × 256 pixels before importing to the network. Network models were trained to keep the quality of the output image close to that of the image from the input image with the number of pooling and upsampling. The training set was used to adjust the parameters of DCNN, and the test set was used to assess the performance of the model. The performance of the DCNN was evaluated using a fivefold cross-validation procedure. Dice similarity coefficient (DSC) was calculated to evaluate the detection accuracy between manual contour and auto segmentation. The DCNN was trained in approximately 17 min with a time step of 20 s/epoch using a graphics processing unit. Once the model had been trained, approximately 30 msec per one image were required to detect the intestinal gas from a new kV image. The training and validation accuracy of the models after 50 epochs calculating were 0.94 and 0.85, respectively. Average±standard deviation DSC for 30 test images was 0.84±0.08. The proposed DCNN method can automatically detect the intestinal/rectal gas in kV images with good accuracy. It could be useful to detect the intestinal/rectal gas in prostate cancer patient.