DeepRegYLab: A Dilated Convolution and Network Design Space Based Image Segmentation Model to Segment Healthy GI Tract of Cancer Patients to Automate Radiotherapy

Gastrointestinal (GI) tract cancer is the most preva-lent type of cancer globally. It's crucial to distinguish between healthy and diseased cells before administering radiation treatment to cancer cells. To segment healthy GI tract organs for proper radiation therapy, we proposed DeepRegYLab, a dilated convolution, and a network design space-based image segmentation model. A RegNetY backbone, an encoder, and a CNN-based decoder were employed in the proposed model. The encoder consists of dilated convolution, 1 x 1 convolution, and an upsampling operation. RegNetY is a network design space containing convolution operation, SE block, batch normalization, and activation function. On the G I tract MRI scan dataset, our proposed model achieved good accuracy and minimal validation loss. The dice score, the IOU value, and the validation loss of our proposed model were 0.83, 0.77, and 0.12 respectively. The proposed model outperformed many other image segmentation models when it comes to segmenting healthy G I tract organs from MRI 2D scans.