Towards reliable no-reference image quality assessment in photoacoustic imaging

Assessing image quality in photoacoustic imaging remains an open challenge, particularly due to the lack of reference images in in vivo settings. While simulated data can provide ground truth for objective evaluation, traditional metrics often fail to capture artefacts and degradations that affect perceptual quality. This limitation applies to both in vivo and synthetic data, underscoring the need for perceptually meaningful no-reference metrics. This work explores a no-reference image quality metric for photoacoustic imaging by adapting an existing deep learning framework developed to assess image quality as perceived by humans. The model was trained and tested on simulated forearm images with varying detector geometries and added noise levels, creating a quality hierarchy based on a score that accounts for the severity of limited-view artefacts and the peak signal-to-noise ratio (PSNR) of Delay-and-Sum (DAS) reconstructions. Preliminary results show that the model effectively captures the considered degradations, achieving an R² score of 0.83 in predicting the defined image quality rankings and performing better than traditional metrics such as contrast-to-noise ratio (CNR).