A radiomics-based machine learning framework for full-reference ultrasound image quality assessment across varying acquisition settings.

BACKGROUND AND

OBJECTIVE

Ultrasound image quality depends on acquisition conditions, while conventional full-reference image quality assessment (FR-IQA) metrics may not fully capture perceptually relevant intensity and texture changes. Radiomics has been explored for image-quality characterization in other medical modalities. This study developed a radiomics-based machine-learning framework for FR-IQA under controlled acquisition-setting variations.

METHODS

Rabbit cadaver ultrasound data included two anatomies and two probe-scanner configurations, defining four reference groups. Varying frequency, gain, power, and dynamic range yielded 192 reference-test pairs. Four sonographers rated pairs. B-mode and wavelet radiomic descriptors formed six full-reference representations evaluated by reference-grouped nested model development. Performance was compared with eight classical FR-IQA methods and two learned perceptual metrics, with robustness and acquisition-setting analyses.

RESULTS

Average-measure inter-observer agreement was high, with ICC(A,k) = 0.9602. Out-of-reference predictions achieved SROCC = 0.8930, PLCC = 0.8702, RMSE = 0.1462, and R2 = 0.7196. On the primary region of interest (ROI), the framework showed higher observed aggregate performance than all classical FR-IQA methods, while learned perceptual metrics showed higher observed performance. No reference-aware pairwise comparison was significant after Holm correction. Predictions were stable under modest ROI and discretization perturbations. Gain had the largest median radiomic effect and was the dominant factor for subjective quality and predictions across reference groups, enabling feature- and prediction-level assessment of acquisition-setting sensitivity.

CONCLUSIONS

Radiomic full-reference representations provide an interpretable approach for characterizing acquisition-dependent ultrasound image quality, with potential utility for system evaluation and protocol optimization. Larger multi-system, in vivo studies with repeated acquisitions and independently defined references are needed to establish generalizability.