AI Development on Rectal Cancer by using Medical Imaging.

Colorectal cancer (CRC) is a major global health issue and a leading cause of cancer-related deaths. Early and accurate diagnosis is crucial for improving patient outcomes, yet current meth- ods for analyzing MRI scans—a key tool for tumor detection and staging—often rely on manual segmentation, a process that is time-consuming and prone to variability between observers. This thesis presents a new, AI-driven framework that integrates advanced deep learning techniques with radiomics analysis to enhance the identification and characterization of CRC tumors from MRI scans. We introduce and evaluate three innovative deep learning models, 2D-UNet, 2D-UMamba, and 2D-Swin-UMamba, that automatically segment tumors on MRI images. The models are trained and tested on a carefully curated dataset of rectal cancer patients, ensuring robust performance across diverse imaging conditions. After segmentation, radiomics is used to extract quantitative features that describe tu- mor shape, texture, and intensity. These features are analyzed for correlations with clin- ical factors such as tumor stage, microsatellite instability (MSI), and genetic mutations (KRAS/NRAS/BRAF), offering a more comprehensive and quantitative assessment of CRC tumors. This combined approach has the potential to support more personalized treatment plan- ning and improved patient care. Our experimental results demonstrate that the proposed AI models outperform traditional segmentation methods, achieving higher accuracy as measured by the Dice Similarity Coeffi- cient (DSC) and Intersection over Union (IoU). Furthermore, the radiomics analysis indicates that AI-driven feature extraction can identify imaging biomarkers predictive of clinical out- comes, potentially reducing the need for invasive diagnostic procedures. In general, this re- search contributes to a scalable and clinically relevant framework for automated MRI-based segmentation and radiomic analysis in the management of colorectal cancer.