Assessing image quality in photoacoustic imaging: A metric-based and deep learning-based evaluation.

Area of Science:

• Biomedical Imaging
• Medical Physics
• Image Processing

Background:

• Photoacoustic (PA) imaging provides high-resolution functional *in vivo* imaging, but its image quality is sensitive to acquisition hardware, reconstruction algorithms, and scanning parameters.
• Consistent and reliable image quality assessment (IQA) is crucial for advancing PA imaging technologies and ensuring research reproducibility.
• Existing IQA metrics, often developed for natural images, lack systematic validation for the unique characteristics of PA imaging data.

Purpose of the Study:

• To conduct the first large-scale benchmark of IQA metrics specifically for PA imaging.
• To evaluate the performance of full-reference (FR) and no-reference (NR) IQA metrics across diverse PA image datasets.
• To develop and validate deep learning models for automated, reference-free PA image quality estimation.

Main Methods:

• Evaluated 11 FR and 2 NR IQA metrics on nearly one million PA images from five distinct datasets (phantoms, preclinical, *ex vivo*, *in vivo*).
• Acquisitions included controlled degradations across multiple commercial PA imaging systems.
• Trained three deep learning architectures (PAQNet, IQDCNN, EfficientNetIQA) to predict metric scores directly from PA images for automated NR assessment.

Main Results:

• Structural similarity-based FR metrics, particularly Structural Similarity Index Measure (SSIM) and its variants, demonstrated consistent ability to differentiate PA image quality.
• Conventional metrics like Peak Signal-to-Noise Ratio (PSNR) and existing NR metrics showed poor correlation with reconstruction improvements.
• Deep learning models, especially PAQNet, achieved high correlation with reference-based scores, offering a practical approach to automated quality assessment, though cross-system generalization remains a challenge.

Conclusions:

• This study establishes a benchmark for PA image quality evaluation, highlighting the superiority of structure-aware metrics.
• Learned NR predictors, like PAQNet, enable more reliable and automated quality assessment in PA imaging.
• The findings support the development of robust IQA tools for advancing PA imaging research and clinical translation.