Quality assessment of optical coherence tomography angiography images with Relative-distance-based Patch Distribution...

Optical coherence tomography angiography (OCTA) is a non-invasive technique to visualize changes in retinal microvasculature. The relationship between the retinal blood flow and systemic diseases can be studied using biomarkers extracted from the OCTA en-face images. However, these images are prone to various artifacts that can hinder the image quality. Thus, image quality assessment is essential to improve the retinal biomarker analysis pipeline.

In this paper, we propose a novel method called Relative-distance-based Patch Distribution Modeling (R-PaDiM) that compares the probabilistic representation of good and bad quality images in a relative manner using patch features extracted from pre-trained encoders. With our method, it is possible to both classify OCTA en-face images into good and bad quality, and obtain patch-wise quality score maps to highlight the bad quality regions within the image for better interpretability. Five different backbones are thoroughly investigated for image quality assessment on two public and one private datasets: DRAC Challenge, OCTA-25K-IQA-SEG, and MeyeHeart.

We achieve state-of-the-art results on all backbones and datasets. Our best results are observed on the DRAC Challenge dataset with a WideResNet-50 backbone that has an accuracy of 98.0 ± 1.1, an AUC of 99.3 ± 0.5, and a Kappa score of 86.4 ± 7.6. We also report a higher correlation between the patch-wise quality scores and the artifact affected regions compared to other methods.

The proposed method is highly robust and efficient in obtaining quality scores that correspond to the specific regions within the image related to the classification decision. Due to its versatile nature, it can be applied to many other tasks with various imaging modalities for better explainability of the deep learning models.