Self-supervised learning for automated image quality assessment in endoscopy

Computer-aided detection and diagnosis systems (CADe/x) in endoscopy are primarily trained on high-quality imagery from expert centers. However, various factors can significantly impact the image quality of endoscopic images during endoscopic procedures in clinical practice. This discrepancy between the high-quality conditions of training and validation and the more heterogeneous conditions encountered during real-world clinical use could potentially result in unreliable CADe/x systems, which underscores the need to assess image quality prior to CADe/x processing. Therefore, we conduct a pioneering study focusing on automated Image Quality Assessment (IQA) for endoscopic images. To facilitate this assessment, a novel dataset has been curated for endoscopic IQA, including 5,396 frames from 271 patients. As a method, we propose a self-supervised learning framework to improve performance over currently existing IQA models. Given these insights, our research is a first step towards developing an IQA algorithm capable of precisely identifying and potentially eliminating lowquality endoscopic images for further CADe/x processing. Furthermore, such an algorithm can provide valuable feedback to clinicians, increasing the reliability of CADe/x systems by defining a quality standard.