SACM - South Korea

Permanent URI for this collectionhttps://drepo.sdl.edu.sa/handle/20.500.14154/9665

Browse

Now showing 1 - 1 of 1

Restricted
Improving Feature Selection in Medical Image Segmentation
(Saudi Digital Library, 2025) ALABDULWAHAB, Abrar Sami S; Sang, Woong Lee
Colorectal cancer is considered one of the most common cancers worldwide, representing about one in 10 cancer cases and deaths globally. It starts as small, benign polyps which may turn into cancer. Early detection and removal of polyps is crucial to prevent colorectal cancer and ensure appropriate patient treatment. Due to the polyp features, accurately segmenting it can be challenging. Deep learning methods have been used to detect colorectal polyps by extracting the features. However, most of these approaches have limitations in handling polyp variations and often struggle with generalization when trained on small datasets or when encountering polyps with indistinct boundaries. Therefore, Duck-Net was proposed to segment polyps in colonoscopy images and address these challenges through its architecture, by creating a custom convolutional block and applying a secondary downsampling. However, Duck-Net has some limitations when it comes to polyps that have the same color as the colon, making it challenging for the model to detect these polyps. Therefore, Duck-Net performance needs further enhancement to segment and detect small-size, flat polyps, polyps with unclear edges, and subtle abnormalities, which are clinically significant for proper diagnosis. Attention mechanism, and Conv2DTranspose layer could be used to overcome such problems. Therefore, this thesis proposes a method based on a Duck-Net, integrated with the convolutional block attention module and conv2DTranspse to enhance feature representation, improve interpretability, generate higher-resolution outputs and the ability to capture vital small information from images consistently. This study confirmed that Duck-Net’s performance, when integrated with the convolutional block attention module block and conv2DTranspose layer, further enhanced image segmentation and outperformed the standard method in image segmentation and detection of polyps. It is feasible to segment and detect undetectable small-size, flat-shaped lesion polyps, and polyps with indistinct boundaries, which are considered factors for increased miss rate of colorectal cancer polyp detections.
10 0