Deep learning based MRI-PDFF reconstruction from 3D US for accessible liver steatosis staging

Metabolic dysfunction-associated steatotic liver disease is increasingly prevalent, creating a demand for accessible, non-invasive liver-fat quantification. Magnetic resonance imaging-derived proton density fat fraction (MRI-PDFF) provides a quantitative reference standard for hepatic steatosis, but its cost, limited availability, and examination time restrict routine screening and monitoring. Ultrasound is widely available and inexpensive, yet conventional imaging is qualitative, while quantitative ultrasound (QUS) often relies on handcrafted acoustic biomarkers and system-specific calibration. This thesis investigates whether MRI-PDFF can be estimated directly from volumetric ultrasound radiofrequency (RF) data. Same-day three-dimensional ultrasound and abdominal MRI examinations from 36 participants were spatially paired through clinically reviewed rigid registration, resampling, scan conversion, and liver-mask preparation. This produced 248 aligned RF, MRI-PDFF, and mask triplets. Automatic registration methods were also evaluated as reproducible quality checks. A combined modality independent neighbourhood descriptor (MIND) and Dice objective was most robust to imposed misalignment, although it did not outperform the manual registrations. A conventional reference-phantom attenuation coefficient estimation (ACE) pipeline was implemented as a QUS baseline. Subject-level ACE was strongly associated with MRI-PDFF (Pearson r = 0.827, R² = 0.683), confirming that the RF data contained measurable steatosis-related information. An end-to-end deep learning framework was subsequently developed using overlapping axial RF windows and complementary waveform, envelope, spectral, and depth-dependent representations. A one-dimensional convolutional encoder preserved local signal information, while a two-dimensional dual-decoder U-Net jointly predicted spatial PDFF and liver-validity maps. Using subject-grouped five-fold cross-validation, the final model achieved r = 0.8624 and R² = 0.7437, exceeding the ACE baseline while maintaining stable predictions across repeated acquisitions. Within the MRI-PDFF range of 0–21.1%, performance improved to r = 0.909 and R² = 0.827. Detection of the presence of steatosis was strong (AUROC = 0.980), but sensitivity declined at the moderate and severe steatosis thresholds. These results demonstrate that spatially aware learning from raw ultrasound RF data can estimate MRI-PDFF more accurately than a single handcrafted attenuation biomarker and can produce spatially resolved liver-fat maps. However, predictions were systematically underestimated at high PDFF, reflecting limited severe-disease representation and ultrasound saturation. Larger, independent, multi-system studies are required before clinical translation.