| Author: | Li, Ka Chun |
| Title: | Deep learning-based non-invasive histological grading of hepatocellular carcinoma via multi-modal fusion of dual-tracer PET/CT and clinical data |
| Advisors: | Lin, Liang-ting (HTI) Yoo, Jung Sun (HTI) |
| Degree: | DHSc |
| Year: | 2025 |
| Department: | Faculty of Health and Social Sciences |
| Pages: | xi, 206 pages : color illustrations |
| Language: | English |
| Abstract: | Background Hepatocellular carcinoma (HCC), the main type of primary liver cancer, presents a major challenge in clinical practice due to its increased global prevalence and high mortality rate. Accurate grading of HCC, which shows tumor differentiation from well-differentiated to poorly differentiated, is crucial for predicting prognosis and guiding treatment choice. Traditionally, biopsy has been the primary method for HCC grading, providing direct histological evidence, but it carries significant risks and limitations, such as complications. Researchers highlighted the urgent need to shift from invasive biopsy to non-invasive approaches that utilize advanced imaging models. Among these, positron emission tomography/computed tomography (PET/CT) with dual tracer, ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) and ¹¹C-acetate (¹¹C-ACT), is notable for offering metabolic insight. Meanwhile, deep learning (DL) predictive model provides the possibility to integrate multi-modal data to predict pathological grading of HCC accurately. Method This retrospective study analyzed a cohort of 93 patients diagnosed with HCC who underwent surgical resection or biopsy at the Department of Nuclear Medicine, Hong Kong Sanatorium and Hospital, between January 2004 and December 2024. Inclusion criteria include histopathologically confirmed HCC, dual-tracer PET/CT imaging (¹⁸F-FDG and ¹¹C-ACT) performed within one month of surgery or biopsy, and availability of histopathology report. Dual-tracer PET/CT image (¹⁸F-FDG and ¹¹C-ACT) with clinical variables like age, sex, hepatitis status, and alpha-fetoprotein (AFP) level and histopathology report were collected. Histopathology report served as reference standard pathological grade. Experienced pathologists determined grade (well-differentiated, moderately differentiated, poorly differentiated) based on World Health Organization (WHO) system criteria. Grade used as ground truth for model training and validation. The calibrated imaging and clinical branch feature vectors obtained after processing through 3D SwinUNETR encoder, are fused to create unified imaging feature representation and then passed to two-layer classifier to predict the grading. Four different combinations of Dual-tracer PET/CT image (¹⁸F-FDG and ¹¹C-ACT) and clinical variables for the feature branch were evaluated. Combination of two loss functions and divergence measures are tested: Cross Entropy (CE) and Focal Loss (FC), Jensen-Shannon (JS) Divergence and Kullback-Leibler (KL) Divergence. Five different fusion techniques are evaluated, None, Concatenation, SE (Squeeze-and-Excitation), MHA (Multi-Head Attention), and the proposed EGCA (Early Gated Cross-Attention). This study employs a three-fold cross-validation method. Four key metrics are used to evaluate model performance: Precision, Recall, F1-Score, and Area Under the Curve (AUC). Result The model integrating ¹⁸F-FDG and ¹¹C-ACT PET/CT imaging with clinical features, employing loss function of FC+JS and fusion strategy of proposed EGCA, achieved the best performance, yielding AUC of 0.8034 ± 0.0136 and a precision of 0.7088 ± 0.0501. While for ⁸F-FDG and ¹¹C-ACT PET/CT PET/CT imaging without clinical data achieved AUC of 0.7527 ± 0.0349, precision: 0.6563 ± 0.0240. Integrating clinical features with ACT imaging alone resulted in an AUC of 0.6736 ± 0.0323 (precision: 0.6054 ± 0.0156), while combining FDG PET/CT with clinical features but without ACT imaging produced the lowest performance, with an AUC of 0.6381 ± 0.0248 and a precision of 0.6018 ± 0.0174. The full combination achieved overall highest accuracy in reflecting ability to distinguish full spectrum HCC grade, stemming from synergistic interaction of tracers. Conclusion This study has demonstrated a promising result (AUC of 0.8034) through integrating dual-tracer ¹⁸F-FDG and ¹¹C-ACT PET/CT imaging with clinical data, using a deep learning model with a 3D SwinUNETR encoder, an FC+JS loss function, and the proposed EGCA fusion strategy to predict the pathological grade of HCC in a non-invasive approach, thereby reducing the risk and necessity of invasive biopsy. |
| Rights: | All rights reserved |
| Access: | restricted access |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 8652.pdf | For All Users (off-campus access for PolyU Staff & Students only) | 2.7 MB | Adobe PDF | View/Open |
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/200/14199