Combining Quantitative MRI Measures to Determine Possibility of Enhanced 3D Liver Segmentation Accuracy Compared to Standard Methods

Sammanfattning

Metabolic-associated fatty liver disease (MAFLD) is one of the most common liver diseases worldwide, affecting approximately 24% of the global population. MAFLD can progress to more severe conditions, such as metabolic dysfunction-associated steatohepatitis (MASH) and hepatic fibrosis, with MASH and fibrosis significantly increasing the risk of liver-related mortality. Currently, the gold standard for assessing and diagnosing hepatic fibrosis and MAFLD is liver biopsy, a procedure associated with high costs, invasiveness, and sampling errors. As a non-invasive alternative, quantitative MRI combined with automatic liver segmentation offers valuable imaging biomarkers, such as proton density fat fraction (PDFF) and T1 relaxation time, which provide insights into liver fat content and inflammatory activity. This master’s thesis investigates two methods for automatic liver segmentation from quantitative MRI volumes: multi-atlas segmentation and seeded region growing. These methods are evaluated and compared using Dice score, Hausdorff distance, and the false positive to false negative (FP/FN) ratio. Additionally, the impact of different quantitative maps on segmentation accuracy is tested. To investigate the separability of liver and surrounding tissue, quantitative tissue properties are visualized in a parametric space. The results show that liver and surrounding tissues overlap in all parametricdimensions, although the degree of overlap varies. The smallest overlap was seen in R1 water. This quantitative map also demonstrated increased performance in seeded region growing. The quantitative maps with smaller overlap seem to have the largest effect on the final segmentation accuracy. The results also demonstrate that seeded region growing, compared to multi-atlas segmentation, results in a lower variance between the test patients, as well as higher Dice scores and lower Hausdorff distance, while maintaining a low FP/FN ratio. In conclusion, liver tissue cannot be fully separated from surrounding tissue using parametric mapping alone. Among the tested maps, R1 water contributed the most to segmentation accuracy. The seeded region growing algorithm outperformed multi-atlas segmentation in this thesis.