Dual-Stream Cross-Fusion Learning For Retinopathy of Prematurity Fundus Image Quality Assessment

The considerable variability in images of retinopathy of prematurity (ROP) has the potential to influence the accuracy of clinical analysis and diagnosis. It is therefore essential to ensure the reliability of the assessment of fundus image quality in order to facilitate accurate subsequent tasks. The current automated image quality assessment techniques may not fully encompass the nuanced intricacies of ROP images, potentially resulting in erroneous outcomes. This paper presents a novel, specifically designed approach, namely RIQA, for the assessment of the quality of ROP images. The proposed approach introduces a dual-stream cross-fusion learning method, which is capable of extracting and fusing key ROP grading features, such as blood vessels or ridges, so as to enhance the performance of image quality assessment. In contrast to existing methodologies that merely yield binary classifications (i.e., good or bad), the proposed approach is capable of categorizing images into three distinct classes: 'good', 'usable' and 'unusable'. To be specific, RIQA independently extracts degraded features and structural details, followed by an integration of high-quality features with low-quality features. This process facilitates the classification of images into one of three quality categories. The experimental results conducted on an ROP dataset comprising 4,228 images indicate that RIQA outperforms existing image quality assessment methods by a significant margin, with an AUC value of 85.45%. The proposed framework may potentially be a standard tool for the evaluation of the quality of ROP fundus images.Clinical relevance- Image quality is crucial for accurate diagnosis of retinopathy of prematurity, and ensuring highquality images improves diagnostic precision. Additionally, the RIQA method offers a more refined classification (such as "usable" and "unusable"), helping doctors better filter and utilize images for clinical analysis.