Deep Learning Models for Medical Image Segmentation: Approaches and Applications

Abstract

The increased availability of medical image data, particularly from modalities such as Magnetic Resonance Imaging (MRI), has opened up possibilities to develop smart systems to aid clinical diagnosis and treatment planning. However, the complexity and volume of such data render manual analysis cumbersome. This thesis addresses the problem of automatic stroke lesion segmentation in brain MRI images through deep learning techniques. A thorough experimental evaluation was conducted with three CNN architectures: U-Net, Attention U-Net, and UNet++. The networks were trained and evaluated on a sub-dataset of the ATLAS 2.0 dataset with annotated post-stroke MRI images. The performance of each model was evaluated by Dice coefficient and Intersection over Union (IoU) metrics. The result shows that the standard U-Net worked well for segmentation ( 74.76% ), and the Attention U-Net and UNet++ provided improvements in lesion boundary and location refinement. The findings show the effectiveness of CNN-based models in effectively segmenting stroke lesions from 2D medical images. This study adds value to the medical image analysis field by providing insight into the capabilities of deep learning models for stroke segmentation and acts as a baseline for further improvement using 3D architectures or transformer-based models.