Machine Learning-Based Inpainting Method to Reconstruct Missing or Corrupted Regions in Retinal Images

In this paper, the machine learning-based inpainting method to reconstruct missing or corrupted regions in retinal images, focusing on accurately erasing blood vessels and possibly other structures, is developed. In addition, the article provides a comparative analysis of approaches to reconstruct missing or corrupted regions in retinal images, focusing on accurately erasing blood vessels and other structures. Initial approaches using deep learning convolution detectors and edge detection as additional loss functions introduced significant artifacts and inconsistencies. However, the breakthrough came with the use of a modified masked Gaussian blur technique, which substantially improved the inpainting quality. While minor issues remain due to retinal mask detection inaccuracies, the results are promising for future applications in medical imaging. Improved inpainting resulted in significantly better inpainting quality. The issues with previous methods were largely mitigated, though some problems persisted due to inaccuracies in the retinal mask detection. Sources of inpainting issues were pixels left off on edges of retinal areas as inpainting takes them as source pixels.