Abstract
BACKGROUND: Accurate detection of lesions in endoscopic videos is critical for the diagnosis and treatment of gastrointestinal diseases; however, poor image quality often affects accurate detection. Therefore, we aimed to introduce a new image quality assessment framework that does not require comparison to a reference image. METHODS: Our lesion detection model was trained and validated using 6,228 still endoscopic images retrospectively collected from 867 patients. We analyzed sequences of at least 10 consecutive frames showing detected lesions, known as tubelets, as the core unit of evaluation. These tubelets were classified into fast or slow movements based on movement speed. The model was tested using 100 colonoscopy videos, comparing its performance when trained on randomly selected images with its performance when trained on images selected using our quality assessment method. RESULTS: Quality-Sorted models outperformed Random models. Quality-Sorted models achieved a positive predictive value of 76.2% and 80.9% in fast and slow motion, respectively. CONCLUSIONS: These results underscore the benefit of integrating quality assessment and preprocessing to enhance lesion detection accuracy. Our automated quality assessment and tubelet-based method enhance the reliability and accuracy of DL-based lesion detection. This comprehensive approach improves diagnostic performance and provides a robust basis for clinical decision support in colonoscopy. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12880-025-01993-7.