Thesis

Novel image processing and deep learning methods for liver cancer delineation from CT data

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2025
Thesis identifier
  • T17363
Person Identifier (Local)
  • 201860454
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • One of the important significances of medical image segmentation is its key role in personalized medicine and precision treatment. With the advancement of medical imaging technology, three-dimensional visualization and quantitative analysis of the liver and its lesion areas have been achieved. By accurately segmenting and identifying anatomical structures and lesion areas, the processed images can provide highresolution and quantitative anatomical information to help clinicians develop more effective and personalized treatment plans. In liver surgery or tumour radiotherapy, accurate image segmentation can help plan surgical paths or ensure that drugs are accurately concentrated in the tumour area while avoiding healthy tissues and organs at risk (OARs), thereby improving treatment efficacy and reducing side effects. In the traditional liver segmentation process, radiologists need to manually depict the liver contour, abnormal lesions in the liver (such as cancerous tumour areas), and other key anatomical structures. However, this manual depiction process is easily affected by differences between observers and the same observer at different times, which may affect the accuracy of treatment. Therefore, with the help of advanced imaging and computing technologies, it is of practical significance and value to achieve automated and high-precision segmentation of the liver and liver cancer areas in clinical applications. This thesis designed a series of innovative deep learning methods to achieve automatic segmentation of the liver and liver cancer areas in computed tomography (CT) images, focusing on improving segmentation accuracy, robustness, and the ability to handle small tumour areas. The datasets used in this study are from the public liver segmentation dataset, The Liver Tumour Segmentation (LiTS). In view of the liver CT segmentation problem, this study proposed an improved UNet model based on multi-scale feature fusion, which significantly improved the segmentation performance. The model combines dilated convolution and pyramid pooling modules to enhance the ability to capture multi-scale features. Among them, dilated convolution improves the model's sensitivity to fine-grained features without increasing the number of parameters by expanding the receptive field; the pyramid pooling module enhances the recognition ability of complex liver anatomical structures by aggregating multi-scale global features. Experimental results show that the multiscale feature fusion U-Net performs well in the liver segmentation task, with an average Dice coefficient of 0.95, and still shows stability and robustness when dealing with cases with blurred boundaries. In the liver cancer segmentation task, this study proposed two cascaded U-Net networks based on the attention mechanism, which solved the challenges in this field with different strategies. The first method introduces a custom attention mechanism module based on the cascaded U-Net to simulate the characteristics of humans automatically focusing on cancerous areas during the segmentation process, which significantly improves the accuracy of segmentation. The model can automatically focus on the key features of the tumour in the decoding stage while suppressing the interference of background noise, providing effective support for accurate segmentation. The second method proposes an end-to-end trained cascade hybrid attention network, whose main goal is to directly complete the liver cancer segmentation task through a single network, avoiding the process of using two independent networks in traditional methods. The network achieves precise focus on the liver region by adding a Bounding Box module based on the hard attention mechanism between two U-Nets. At the same time, the channel attention module and the spatial attention module are introduced in the feature extraction stage, so that the model can effectively capture the key features of the tumour region and further improve the segmentation performance. The improved U-Net model performs well in dealing with tumour regions with complex shapes and irregular boundaries. Experimental data show that both cascade U-Nets based on the attention mechanism show excellent performance in the liver cancer segmentation task, maintaining high accuracy even in the case of small tumours and blurred boundaries. The average Dice coefficient of the first method reaches 0.78, while the second method is further optimised to 0.80, which is significantly better than the traditional U-Net of 0.70. In addition, this study developed a novel boundary optimisation method to improve the accuracy of the segmentation boundary by extracting the boundary area of the initial segmentation result and refining it. The optimisation module uses a deep learning-based boundary refinement network (BRN) to fine-tune the boundaries of the liver and tumour. Through block processing, this method effectively reduces false positive and false negative areas and significantly improves the continuity and accuracy of the segmentation results. Multiple case verification results show that the optimised model has improved boundary-related indicators (such as average symmetric distance and boundary overlap error) by about 10%.
Advisor / supervisor
  • Di Caterina, Gaetano
  • Soraghan, J. J.
Resource Type
DOI

Relations

Items