Abstract
Prostate cancer (PCa) is one of the most common cancers and a leading cause of cancer-related deaths among men worldwide. However, the five-year survivor rate of PCa exceeds 97% with early diagnosis and treatment. Multimodality magnetic resonance imaging (MRI) is increasingly used for diagnosis and treatment planning of PCa due to its superior ability to reflect prostate characteristics from various aspects. Specifically, T2-weighted (T2W) MRI is typically used for understanding the prostate anatomy due to its high resolution, while Diffusion-Weighted Imaging (DWI) provides functional information. Although multimodality MRI is a powerful non-invasive technique for detecting PCa, and MRI-guided biopsies can significantly improve PCa diagnosis accuracy while reducing unnecessary pain to patients, the interpretation of multimodality MRI is time-consuming and subject to significant inter-observer variability due to differences in radiologist expertise.Therefore, there is a critical need for computer-aided tools to improve the efficiency and accuracy of prostate MRI interpretation in clinical practice. This thesis develops automatic and efficient human interactive deep learning solutions to assist radiologists in enhancing the efficiency and accuracy of multimodality MRI interpretation in the prostate cancer diagnosis workflow.
In the first part, we propose a boundary-aware semantic clustering network (BASC-Net) to automate the prostate zonal segmentation. Since PCa exhibits different characteristics in different prostate zones and the diagnostic workflows vary accordingly, automated prostate zonal segmentation provides crucial information to streamline the PCa diagnosis process. BASC-Net consists of two major components: the semantic clustering attention (SCA) module and the boundary-aware contrastive (BAC) loss. The SCA module implements a self-attention mechanism that extracts feature bases representing discriminative features of the inner body and boundary subregions and constructs attention maps highlighting each subregion. The features extracted from the inner body and boundary subregions of the same zone were integrated by BAC loss, which promotes the similarity of features extracted in the two subregions of the same zone. The BAC loss further promotes the difference between features extracted from different zones. BASC-Net was evaluated on the NCI-ISBI 2013 Challenge and Prostate158 datasets. An inter-dataset evaluation was conducted to evaluate the generalizability of the proposed method. BASC-Net outperformed ten state-of-the-art methods in all three experimental settings.
In the second part, we develop a prostate lesion segmentation workflow that integrates automatic segmentation and human intervention for improved accuracy and efficiency in lesion segmentation. The proposed framework leverages the benefits of the automated segmentation method with automatic quality assessment to enhance segmentation performance while reducing the burden of user intervention. This approach improves cost-effectiveness by enabling radiologists to focus on cases where automated methods underperform. Three key components are involved in the proposed framework: 1) A coarse segmentation U-Net that automatically generates initial segmentation results. 2) A rejection network that estimates the quality of these initial results, flagging "inaccurate'' cases for user intervention. 3) A Segment Anything Model (SAM) that utilizes both manually and automatically generated ROI information to produce fine segmentation results.
We conducted extensive experiments on Prostate158 and PROSTATEx2 datasets. The results demonstrated that our approach achieved substantial improvements in segmentation performance when approximately 20% of images with the lowest quality scores were flagged for human annotation. With only half of the images manually annotated, the final segmentation accuracy is statistically indistinguishable from that achieved with full manual annotation.
In the final part, we propose an efficient interactive segmentation framework for segmenting prostate zones and lesions from 3D multimodality MRI. The proposed method takes the second part one step further by focusing on efficient interactive segmentation on a patient basis while enabling iterative refinement to further improve segmentation performance. This approach allows users to provide click interactions for only a subset of slices within a 3D volume, enabling efficient generation of initial segmentation results. Users can then refer to the automatically estimated quality score for each result and add correction clicks to progressively refine any unsatisfactory results, thereby improving overall segmentation performance. The overall framework works as follows: 1) Click features are derived from click interactions provided by users. 2) The proposed click interaction propagation (CIP) module reasons click features and positions for slices lacking interactions in the feature space, propagating limited clicks across the whole 3D image. 3) Both predicted and human-provided clicks are used to generate initial segmentation results. 4) The proposed focal click refinement (FCR) module automatically estimates the quality of each result and enables radiologists to refine unsatisfactory slices progressively. Extensive experiments demonstrate that our method achieves superior performance with limited click interactions and competitive segmentation performance with at least 20% fewer interactions compared to state-of-the-art interactive segmentation methods.
The methods proposed in this thesis will streamline the PCa diagnosis workflow and enable efficient human-in-the-loop segmentation. These advancements promote a more efficient and accurate PCa segmentation process, potentially enhancing diagnostic accuracy while reducing the workload for radiologists.
| Date of Award | 3 Dec 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Kei Hang Katie CHAN (Supervisor), Kwok Leung CHAN (Supervisor) & Chi Yuen Bernard CHIU (External Co-Supervisor) |