Gastro-Intestinal Tract Image Segmentation using Edge U-Net and U-Net VGG19

In this work, an approach to tackle the problem of image segmentation on MRI scans of cancer patients using deep learning techniques has been proposed. Radio oncologists are challenged with a tedious task of segmenting the stomach and intestine regions manually. Based on the segmented regions, oncologists need to deliver high doses of X-Ray beams to treat the tumor cells while avoiding the healthy intestine and stomach regions. Using deep learning based segmentation techniques this process can be automated, which would lead to better treatment and save a lot of time. This work proposes a methodology using Edge U-Net and U-Net VGG19 to efficiently segment the large bowel, small bowel and stomach regions from the MRI scans of cancer patients. Experiments were conducted on the proposed methodology and results were compared with U-Net, Feature Pyramid Network and U-Net++ with variant encoders such as ResNet50, Xception, EfficientNet-B0, ResNeXT50. Comparison results show that the proposed methodology Edge U-Net and U-Net VGG19 has achieved the highest weighted Dice coefficient and 3D Hausdorff distance value of 0.86194.