OPTIMIZING 3D MEDICAL IMAGE SEGMENTATION MODELS THROUGH ARCHITECTURE TUNING AND QUANTIZATION TECHNIQUES: BALANCING MODEL EFFICIENCY AND ACCURACY

Abstract

Medical image segmentation is a critical task in medical image analysis and patient diagnosis. This thesis investigates the application of the 3D U-Net architecture and its variations, EquiUnet and Att_EquiUnet, for brain tumor segmentation on the BraTS 2020 dataset. A comprehensive evaluation framework was utilized to assess segmentation performance across whole tumor (WT), tumor core (TC), and enhancing tumor (ET) regions. Results demonstrated the robust performance of the baseline 3D U-Net, achieving high accuracy (91.19%) across all tumor regions. EquiUnet did not exhibit significant performance gains over the baseline U-Net. However, Att_EquiUnet, using the CBAM attention module, showed improvements in boundary localization as evidenced by reduced Hausdorff distances. The study also explored the impact of quantization on model size and accuracy. 16-bit quantization emerged as an optimal compromise, achieving a significant reduction in model size (to 25% of the original) while maintaining accuracy and even slightly improving sensitivity in some cases. 8-bit quantization, while further reducing model size (to 6.4%), incurred a more pronounced accuracy loss, raising concerns about its suitability for clinical use. This thesis contributes to the field by offering a comparative study of U-Net variants for 3D image segmentation and highlighting the potential of attention mechanisms and 16-bit quantization for improving model performance and clinical applicability.