Deep-Based Quality Assessment of Medical Images Through Domain Adaptation

Predicting the perceptual quality of medical imagery is crucial for diagnostic workflows, yet annotated data are scarce and quality metrics must remain reliable across acquisition domains. We introduce a lightweight convolutional self-attention architecture that interpolates global quality scores from local perceptual cues while remaining data efficient.

The method leverages unsupervised and semi-supervised domain adaptation to transfer knowledge between limited annotated datasets and new medical imaging modalities. Evaluations on subjective quality benchmarks show that the proposed approach improves generalisation compared with standard reference-free baselines, enabling practical quality monitoring in clinical pipelines.