Document Type

Thesis

Degree Name

Master of Science (MSc)

Department

Physics and Computer Science

Program Name/Specialization

Applied Computing

Faculty/School

Faculty of Science

First Advisor

Dr. Bernard Chiu

Advisor Role

Thesis Supervisor

Abstract

Late gadolinium enhancement cardiovascular magnetic resonance (LGE-CMR) helps visualize myocardial injury and fibrosis; however, heterogeneous image appearance, poorly defined myocardial boundaries, and subtle differences in enhancement patterns can complicate visual interpretation and automated analysis. This thesis addresses two linked tasks in LGE-CMR: myocardial segmentation and patient-level multiclass disease classification.

First, deep-learning configurations were investigated for myocardial segmentation across heterogeneous phase-sensitive inversion recovery (PSIR) and magnitude-reconstructed (MAG) images. The best-performing configuration achieved a mean patient-level Dice score of 0.823 ± 0.050. For disease classification, a multi-cohort dataset including cardiac amyloidosis, dilated cardiomyopathy, hypertrophic cardiomyopathy, ischemic cardiomyopathy, and myocarditis was used. and healthy individuals. Expert myocardial contours and right ventricular insertion points were used to extract conventional texture features together with descriptors representing the transmural and right-ventricle-aligned angular distribution of myocardial signal. Slice-level features were aggregated across apical, mid- ventricular, and basal regions to form patient-level representations. Feature selection and shrinkage-regularized linear discriminant analysis were performed within patient-wise five- fold cross-validation.

Adding the spatial descriptors to the texture-only representation improved classification accuracy from 0.792 ± 0.048 to 0.853 ± 0.033 and the macro-averaged F1-score from 0.794 ± 0.049 to 0.854 ± 0.035. The spatial framework also outperformed the ConvNeXt- Tiny image-based baseline and reduced specific confusions between disease groups with overlapping global texture characteristics. Finally, MAG-derived texture features were combined with the established PSIR spatial representation. Compared with the PSIR spatial model, fusion increased classification accuracy from 0.853 ± 0.033 to 0.867 ± 0.040, corresponding to an increase from 394 to 400 correctly classified patients. It also improved macro-precision from 0.859 ± 0.037 to 0.877 ± 0.037 and macro-recall from 0.855 ± 0.034 to 0.868 ± 0.041.

Overall, the findings show that myocardial texture is more informative when its anatomical and spatial information is preserved. The proposed methods provide a reproducible framework for spatially informed patient-level characterization, while paired MAG reconstruction offers a modest complementary refinement without requiring an additional imaging acquisition.

Convocation Year

2026

Convocation Season

Fall

Download

Available for download on Thursday, March 23, 2028

Share

COinS