Adaptive Liver Stiffness Quantification in Magnetic Resonance Elastography Using a Heterogeneity-Aware Deep Learning Framework.
Authors
Affiliations (3)
Affiliations (3)
- Research Laboratory of Biophysics and Medical Technologies, University of Tunis El Manar, Higher Institute of Medical Technologies of Tunis, 1006, Tunis, Tunisia. [email protected].
- Research Laboratory of Biophysics and Medical Technologies, University of Tunis El Manar, Higher Institute of Medical Technologies of Tunis, 1006, Tunis, Tunisia.
- Department of Radiology, Military Hospital of Tunis, Tunis, Tunisia.
Abstract
Magnetic resonance elastography (MRE) enables non-invasive liver stiffness measurement (LSM). However, accurate quantification remains operator-dependent particularly for identifying clinically valid measurement regions and selecting a representative sampling strategy. This study developed and validated a fully automated, heterogeneity-aware deep learning framework that integrates valid-region identification, elastogram heterogeneity assessment, and a study-specific adaptive quantification strategy, while substantially reducing analysis time. The framework was evaluated in a prospective cohort of 201 subjects. Magnitude images and wave-attenuation images were combined to identify valid regions for LSM. Because spatial stiffness heterogeneity may increase sensitivity to region-of-interest (ROI) placement, elastograms were classified as homogeneous or heterogeneous using a multimodal convolutional neural network-deep neural network (CNN-DNN) classifier that integrates spatial image features with quantitative stiffness descriptors. For homogeneous elastograms, a study-specific distributed circular-ROI reduced-sampling strategy was evaluated, whereas volumetric sampling was applied to heterogeneous elastograms to better capture spatial variations in stiffness. The segmentation module achieved a Dice Similarity Coefficient (DSC) of 0.894 ± 0.035, an Intersection over Union (IoU) of 0.810 ± 0.057, and a 95th-percentile Hausdorff-Distance (HD95) of 5.45 ± 4.05 mm. Automated LSMs showed excellent agreement with expert measurements (ICC = 0.998; Bland-Altman bias = 0.007 kPa), while reducing analysis time from approximately 15 min to 0.35 s/patient. The heterogeneity classifier achieved accuracies of 89.4% and 88.5% at the slice and patient-levels, with areas under the receiver operating characteristic curve (AUCs) of 0.928 and 0.935, respectively, demonstrating substantial patient-level agreement with the expert reference classification (Cohen's κ = 0.770). By integrating automated valid-region identification with objective elastogram heterogeneity assessment, the framework may improve standardization and reproducibility of LSM.