| Author: | Liu, Anran |
| Title: | AI-assisted whole-slide imaging analysis for hepatocellular carcinoma prognosis : development of a risk scoring system with enhanced interpretability and efficiency |
| Advisors: | Cai, Jing (HTI) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Subject: | Liver -- Cancer -- Prognosis Diagnostic imaging -- Data processing Artificial intelligence -- Medical applications Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Health Technology and Informatics |
| Pages: | 110 pages : color illustrations |
| Language: | English |
| Abstract: | Pathological images serve as the gold standard for cancer diagnosis and play a crucial role in clinical practice. With digital scanner development, Whole Slide Image (WSI) has gained attention due to its larger visual field and clearer imaging. Advanced technologies, especially deep learning models, have driven digital pathology development for downstream clinical tasks like automated diagnosis and prognostic analysis. However, existing methods are either limited by WSI's ultra-high resolution, leading to inefficient analysis, or rely on black box approaches lacking interpretability and clinical acceptability. This thesis addresses limited interpretability and excessive computational complexity in existing WSI automated analysis frameworks. Following pathologists' diagnostic steps, we first propose an attention-driven mechanism named Attention Activator for identifying potential high-risk tissues. This system intuitively highlights potential high-risk areas in WSI using attention maps without prior guidance. Based on these localizations, we construct a multi-perspective risk scoring system from micro to macro levels. Unlike existing systems, we leverage deep learning for multi-perspective feature fusion. Validation results from hepatocellular carcinoma patient cohorts demonstrate our scoring system refines existing clinical staging, identifying potential high-risk patients within low-risk groups. Given WSI's ultra-high resolution, we propose a novel encoding strategy termed FuzzyMIL to enhance analysis efficiency. This approach extends Fuzzy c-means to a learnable transformer framework, leveraging soft clustering characteristics to address feature homogenization in global attention during encoding. This improves downstream task prediction accuracy while minimizing computational complexity. Experimental results showed FuzzyMIL outperforms existing state-of-the-art methods in diagnosis and subtyping across three public datasets while significantly reducing model parameters. Furthermore, we applied the deep learning-based WSI analysis framework to explore automated localization, prognostic analysis, and treatment efficacy assessment based on Vessels that Encapsulate Tumor Clusters (VETC). VETC is a recently identified vascular pattern implicated in cancer metastasis progression and patient prognosis. However, current analyses rely on clinical experience, lacking standardized, quantitative methods. In this pioneering work integrating deep learning with VETC analysis, we developed VETC Net, which automatically distinguishes and locates VETC+ and VETC- regions within WSI. This network facilitates automated VETC+ distribution assessment and enables precise quantification of VETC-prognostic risk correlation. We also evaluated VETC response to various treatment modalities. Results demonstrate VETC Net enables precise VETC tissue localization and accurate prognostic assessment across multi-center, multi-treatment datasets. In conclusion, this thesis presents solutions to limited interpretability and high computational complexity challenges in existing WSI analysis frameworks. We introduce Attention Activator for high-risk tissue identification and a multi-perspective risk scoring system improving risk stratification. Additionally, we propose FuzzyMIL, enhancing analysis efficiency while reducing model complexity. We explored novel VETC-based clinical tasks for automated localization and prognostic analysis. Through extensive experimental validation, our methods show significant improvements in interpretability, efficiency, and clinical applicability for WSI analysis. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14178