Exploring Deep Learning-Based Techniques for 3D Medical CT Image Segmentation
Published in , 2024
Abstract: In recent years, novel medical imaging technologies, epitomized by CT imaging, have emerged and been widely used, becoming a significant auxiliary means for clinical diagnosis and treatment. Since the images generated by CT scans are three-dimensional with large data volume, in order to help clinicians locate the lesion area faster and make accurate diagnoses, it is an exigent need for automatic and precise segmentation of key targets within medical images through computer assistance. With the maturation of deep learning technologies and further enhancement of computing power, segmentation methods based on convolutional neural networks have become the mainstream choice in the field of medical image segmentation, due to their potent capacity for context extraction. However, medical CT images often exhibit noise, motion artifacts, and uneven contrast among other issues, which leads to a notable difference from natural images. Some existing segmentation methods designed for natural scenarios often have poor results if directly applied to medical images. Besides, pixel-wise annotation for medical images requires considerable time and manpower costs. It is usually costly to obtain large-scale, high-quality annotated datasets, especially for 3D data. Given the above status quo, 3D medical image segmentation tasks face numerous challenges. This thesis conducts research on exploring deep learning-based techniques for 3D medical CT image segmentation. First, a comprehensive analysis of prevalent 3D medical image segmentation methods is undertaken. Afterwards, replications are carried out on both 3D segmentation models for small-scale datasets (such as nn-UNet and UNETR) and unified 3D segmentation models for large-scale datasets (such as SAMMed3D) respectively, with experiments being performed on three publicly available medical CT image segmentation datasets: the MSD spleen dataset, BTCV dataset, and WORD dataset. Lastly, enhancements for existing 3D medical CT image segmentation methods are proposed.