Hideharu Miura1, Shuichi Ozawa2, Yoshiko Doi2, Minoru Nakao2, Keiichi Ohnishi3, Masahiro Kenjo2, Yasushi Nagata2. 1. Hiroshima High-Precision Radiotherapy Cancer Center, Hiroshima, Japan; Department of Radiation Oncology, Institute of Biomedical & Health Science, Hiroshima University, Hiroshima, Japan. Electronic address: miura@hiprac.jp. 2. Hiroshima High-Precision Radiotherapy Cancer Center, Hiroshima, Japan; Department of Radiation Oncology, Institute of Biomedical & Health Science, Hiroshima University, Hiroshima, Japan. 3. Hiroshima High-Precision Radiotherapy Cancer Center, Hiroshima, Japan.
Abstract
PURPOSE: The detection of intestinal/rectal gas is very important during image-guided radiation therapy (IGRT) of prostate cancer patients because intestinal/rectal gas increases the inter- and intra-fractional prostate motion. We propose a deep convolutional neural network (DCNN) to detect intestinal/rectal gas in the pelvic region. MATERIAL AND METHODS: We selected 300 anterior-posterior kilo-voltage (kV) X-ray images from 30 prostate cancer patients. Thirty images were randomly chosen for a test set, and the remaining 270 images used as the training set. The intestinal/rectal gas was manually delineated on kV X-ray images and segmented. The training images were augmented by applying artificial shifts and fed into a DCNN. The network models were trained to keep the quality of the output image close to the quality of the input image by pooling and upsampling. The training set was used to adjust the parameters of the 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. The dice similarity coefficient (DSC) was calculated to evaluate the detection accuracy between the manual contour and auto-segmentation. RESULTS: The DCNN was trained within approximately 17 min with a time step of 20 s/epoch. The training and validation accuracy of the models after 50epochs were 0.94 and 0.85, respectively. The average ± standard deviation of the DSC for 30 test images was 0.85 ± 0.08. CONCLUSIONS: The proposed DCNN method can automatically detect the intestinal/rectal gas in kV images with good accuracy.
PURPOSE: The detection of intestinal/rectal gas is very important during image-guided radiation therapy (IGRT) of prostate cancerpatients because intestinal/rectal gas increases the inter- and intra-fractional prostate motion. We propose a deep convolutional neural network (DCNN) to detect intestinal/rectal gas in the pelvic region. MATERIAL AND METHODS: We selected 300 anterior-posterior kilo-voltage (kV) X-ray images from 30 prostate cancerpatients. Thirty images were randomly chosen for a test set, and the remaining 270 images used as the training set. The intestinal/rectal gas was manually delineated on kV X-ray images and segmented. The training images were augmented by applying artificial shifts and fed into a DCNN. The network models were trained to keep the quality of the output image close to the quality of the input image by pooling and upsampling. The training set was used to adjust the parameters of the 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. The dice similarity coefficient (DSC) was calculated to evaluate the detection accuracy between the manual contour and auto-segmentation. RESULTS: The DCNN was trained within approximately 17 min with a time step of 20 s/epoch. The training and validation accuracy of the models after 50epochs were 0.94 and 0.85, respectively. The average ± standard deviation of the DSC for 30 test images was 0.85 ± 0.08. CONCLUSIONS: The proposed DCNN method can automatically detect the intestinal/rectal gas in kV images with good accuracy.