A Novelty Patching of Circular Random and Ordered Techniques on Retinal Image to Improve CNN U-Net Performance

U-Net is one of the CNN architectures with deep layers that requires a lot of training and testing data. Unfortunately, the retinal image data is not freely available. The large images can add parameters and network complexity for U-Net. To improve the ability of U-Net in retinal blood vessels segmentation, it is necessary to provide sufficient with small data size. The blood vessels of the retina are located inside the retina circle. This study proposes a circular random patching and ordered patching technique to cut an image in certain size. The circular random patching uses the Euclidean distance at any point inside the retinal circle. If the patching technique is applied carelessly and does not pay attention to the circle area of the retina, it could cause the outer part of the retina circle that does not contain blood vessels to be taken. This technique proposes to meet the training data requirements of the U-Net architecture by generating as possible many patch images as containing retinal blood vessels. The ordered patching technique is used to obtain the patch images by cutting the entire original image orderly starting from the first pixel point in the image at the testing stage. At final testing stage, the output of patch images has to be reassembled to be reconstructed to its original size with segmented retinal blood vessels features. The proposed method has been implemented on the DRIVE and STARE datasets. The results obtained in DRIVE produced excellent accuracy, specificity sensitivity, F1-score, and IoU. The results in STARE were quite well for accuracy, specificity, and sensitivity, but the F1-score and IoU on STARE are insufficient so that both scores are needed to be improved. The research shows that both the circular random patching in the training stage and ordered patching in the testing stage able to improve the performance of U-Net. The next step for future work is going to combine the proposed patching technique with other CNN architectures to improve the results of retinal blood vessels segmentation performance.