Development of Deep Learning with RDA U-Net Network for Bladder Cancer Segmentation

Simple Summary This study proposed the "Residual-Dense-Attention" (RDA) U-Net model architecture to automatically segment organs and lesions in computed tomography (CT) images. The RDA U-Net used ResBlock and DenseBlock at the encoder. Attention gates were used at the decoder position to help the model suppress irrelevant areas of the CT image. Forty-one patients' bladder images were provided for training. The RDA U-Net provided faster but satisfactory segmentation results for bladder cancers and lesions. In today's high-order health examination, imaging examination accounts for a large proportion. Computed tomography (CT), which can detect the whole body, uses X-rays to penetrate the human body to obtain images. Its presentation is a high-resolution black-and-white image composed of gray scales. It is expected to assist doctors in making judgments through deep learning based on the image recognition technology of artificial intelligence. It used CT images to identify the bladder and lesions and then segmented them in the images. The images can achieve high accuracy without using a developer. In this study, the U-Net neural network, commonly used in the medical field, was used to extend the encoder position in combination with the ResBlock in ResNet and the Dense Block in DenseNet, so that the training could maintain the training parameters while reducing the overall identification operation time. The decoder could be used in combination with Attention Gates to suppress the irrelevant areas of the image while paying attention to significant features. Combined with the above algorithm, we proposed a Residual-Dense Attention (RDA) U-Net model, which was used to identify organs and lesions from CT images of abdominal scans. The accuracy (ACC) of using this model for the bladder and its lesions was 96% and 93%, respectively. The values of Intersection over Union (IoU) were 0.9505 and 0.8024, respectively. Average Hausdorff distance (AVGDIST) was as low as 0.02 and 0.12, respectively, and the overall training time was reduced by up to 44% compared with other convolution neural networks.