ADL-UNet: adaptive dilated learning U-Net architecture for advanced colorectal polyp segmentation

Colorectal cancer (CRC) is a leading cause of mortality worldwide. Early detection of precancerous lesions can significantly reduce CRC incidence by allowing for the removal of polyps during colonoscopy or early intervention in cases of detected cancer. Colonoscopy remains the gold standard for identifying these polyps, but smaller lesions can often be missed. Computer-aided detection (CAD) systems assist in enhancing detection accuracy. However, traditional U-Net models struggle with capturing intricate details, especially for smaller polyps. This thesis introduces an advanced U-Net architecture incorporating Dilated Convolution with Learnable Spacings (DCLS) to overcome these limitations. DCLS dynamically adjusts the receptive field based on input data, capturing fine details of minor polyps and long-range dependencies. This approach reduces the risk of overfitting to training data while ensuring fewer number of parameters. ADL-UNet achieved a 7.4% increase in Dice score on validation and a 6.0% improvement on the test set, demonstrating its reliability in clinical applications. Additionally, ADL-UNet has approximately 50% fewer parameters than the baseline U-Net, highlighting its efficiency. These results indicate that the DCLS-enhanced U-Net significantly improves colorectal polyp segmentation performance, advancing CAD systems for early CRC diagnosis and screening, ultimately improving patient outcomes. The complete codebase and trained models are available at https://github.com/MeghaDeySarkar/ADL-Unet.git .