Dissertation: Exploring Multi-Organ Modeling and Segmentation for Enhanced Diagnostic Applications
In the field of medical image analysis (MIA), precise image segmentation faces numerous challenges. Over the past decades, segmentation methods targeting specific objects have emerged, yielding satisfactory results for individual targets within images. However, single-object segmentation fails to effectively leverage the interplay of organs and tissues within biological structures, leading to oversimplified segmentation objectives. Given the intricate anatomical structures of organisms, implementing segmentation methods for multiple targets proves highly challenging, impeding researchers' progress. Existing multi-organ segmentation methods are relatively scarce and labor-intensive. Moreover, MIA encompasses an extensive development process involving multiple complex steps, hindering developers' efforts to enhance algorithm development efficiency. Therefore, accurate segmentation of multiple targets and the construction of efficient development platforms remain crucial issues that warrant further exploration, offering significant implications for clinical-assisted diagnosis.
This thesis expands the research on MIA from three perspectives. These include the construction of multi-organ deformable models, rapid segmentation of mouse CT images based on the multi-organ deformable models, and the development of a universal software for various medical image analyses. Firstly, we explore the deformation capabilities of three-dimensional (3D) multi-organ deformable models. We employ an improved Principal Component Analysis (PCA) method and a neural network-based approach to investigate deformation modes with different regularities. These studies reveal the additional linear and nonlinear deformation capabilities of the deformable models. The results indicate that the improved PCA method, enables fast and accurate modeling of multiple organs with complex structures and limited samples. The constructed models hold significant clinical significance for precise organ segmentation in medical images, with promising prospects for further research and applications in human clinical diagnosis. Additionally, we find that the neural network-based approach can extract more nonlinear deformation modes of organ models compared to the PCA method. These potential nonlinear deformation modes provide valuable references for further optimization of organ models in the future.
Secondly, as deformable models serve as shape prior knowledge, we combine the deformable models constructed using the improved PCA method with human interaction to simultaneously segment multiple organs in mouse CT images. The segmentation results demonstrate that this technique enables fast and accurate segmentation of multiple organs, with improved robustness.
Finally, we develop AnatomySketch (AS), a software that efficiently implements MIA algorithms, including deformable model construction and multi-organ segmentation, among other functions. The software features a flexible plugin interface and a user-friendly GUI, facilitating efficient implementation and visualization of deformable models and segmentation results. The results show that AS software bridges the gap between laboratory prototyping and clinical testing, accelerating the development of MIA algorithms. In conclusion, our work successfully achieves multi-resolution, multi-organ modeling, and multi-organ segmentation. Furthermore, the developed AS software exhibits powerful capabilities in assisting the development of MIA algorithms.
In this thesis, the modeling and segmentation of multiple organs in abdominal CT images of mice demonstrate that deformable models with a greater variety of deformation modes can provide more valuable prior knowledge. Enriched prior knowledge contributes to more accurate segmentation of multiple targets in the images. The AS software developed in this study is an open-source tool aimed at different algorithm developers, allowing continuous enhancement and expansion of its functionalities. In future research, we aim to further apply the methods presented in this paper to achieve multi-object segmentation in human medical images. Additionally, we intend to utilize the AS software developed in this study to facilitate the development of various medical image algorithms by researchers. Once the software is further refined, it can assist the modeling and segmentation methods proposed in this thesis, enabling real-time interactive segmentation applications in clinical diagnosis.
More information
Zhonghua Chen defends his doctoral dissertation “Multi-organ Medical Image Analysis, Modeling, and Segmentation Exploiting Pre-existing Knowledge”. 16 February 2024 at 12:00.
Opponent is Associate Professor Gongning Luo(Harbin Institute of Technology, China) and custos is Professor Lauri Kettunen (Ģֱ). The language of the dissertation is English. The dissertation can be followed in the lecture hall or online.
.
