Haoli Yin, Rachel Eimen, Daniel Moyer, and Audrey K. Bowden. “SpecReFlow: An Algorithm for Specular Reflection Restoration Using Flow-guided Video Completion.” Journal of Medical Imaging, vol. 11, no. 2, 024012, April 2024.
Specular reflections (SRs) are bright spots in endoscopy videos that can hinder a surgeon’s view and decision-making. Existing methods to remove these artifacts are often slow and prone to errors. SpecReFlow is introduced as the first comprehensive deep-learning solution to detect and restore SR areas in endoscopy videos, maintaining both spatial and temporal consistency.
SpecReFlow operates in three stages: first, an image preprocessing stage enhances the video contrast; second, a detection stage identifies where SR regions are located; third, a restoration stage replaces the SR pixels with accurate representations of the underlying tissue using optical flow to blend color and structure from adjacent video frames.
Tests show that SpecReFlow outperforms previous methods. The detection stage achieves a Dice score of 82.8% and a sensitivity of 94.6%. The restoration stage effectively combines information from multiple frames, providing more accurate restorations than methods using single frames.
SpecReFlow uniquely integrates temporal and spatial data for SR detection and restoration, outperforming older techniques that rely solely on spatial data from single frames. The software is designed for easy deployment in clinical settings, enhancing endoscopy video quality to support accurate diagnosis and treatment. Future improvements will focus on real-time application optimization.
