Stack-U-Net: refinement network for improved optic disc and cup image segmentation

Abstract

In this work, we propose a special cascade network for image segmentation, which is based on the U-Net networks as building blocks and the idea of the iterative refinement. The model was mainly applied to achieve higher recognition quality for the task of finding borders of the optic disc and cup, which are relevant to the presence of glaucoma. Compared to a single U-Net and the state-of-the-art methods for the investigated tasks, the presented method outperforms others by multiple benchmarks without a need for increasing the volume of datasets. Our experiments include comparison with the best-known methods on publicly available databases DRIONS-DB, RIM-ONE v.3, DRISHTI-GS, and evaluation on a private data set collected in collaboration with University of California San Francisco Medical School. The analysis of the architecture details is presented. It is argued that the model can be employed for a broad scope of image segmentation problems of similar nature.

Joint Optic Disc and Dup Segmentation Based on Densely Connected Depthwise Separable Convolution Deep Network

“…From Table 5, We have achieved 2.8% and 9.23% higher than the original U-net in the Dice coefficients of the optic disc and the optic cup, respectively. When compared with the same improved method modified based on the U-net structure [16], [21], [10], [25], our method is better than the best improved method proposed by Shuang Yu [25], which is 0.42% higher in OD, and 2.46% in the more difficult OC segmentation task. On REFUGE dataset, our method is also better than that of M-net [16] using this dataset.…”

“…In order to verify that our method is better than other methods, we compare our proposed method with the state-of-the-art approaches, such as pOSAL framework [19], GL-Net [20], M-Net [16], Stack-U-Net [21], WGAN [22], two-stage Mask R-CNN [23], multi-modal self-supervised pretraining network [24], Shuang Yu [25] and A. Sevastopolsky [10]. Additionally, we compare with the Fully convolutional network U-Net [15].…”

“…The first row is the input image, the second is the ground truth, the third rows is the simple DDSC-Net segmentation results, fourth row is the DDSC-Net without multi input,fifth row is the results of DDSC-Net and sixth row is the reprocess result,where the white region denote the background, and the gray and black region denote the OD and OC segmentations. Tables Table 1 Localization accuracy [15] 0.950 0.820 --M-Net [16] 0.959 0.866 --Stack-U-Net [21] 0.970 0.890 --A. Sevastopolsky [10] -0.850 -0.750 Shuang Yu [25] 0.973 0.887 --MULTI-MODEL-PRE-TRAINING [24] 0.960 0.902 0.924 0.823 WGAN [22] 0.954 0.840 0.912 0.724 pOSAL [22] 0.965 0.858 --GL-Net [20] 0.971 0.905 --Proposed 0.978 0.912 0.957 0.844…”

Joint Optic Disc and Dup Segmentation Based on Densely Connected Depthwise Separable Convolution Deep Network

“…From Table 5, We have achieved 2.8% and 9.23% higher than the original U-net in the Dice coefficients of the optic disc and the optic cup, respectively. When compared with the same improved method modified based on the U-net structure [16], [29], [10], [33], our method is better than the best improved method proposed by Shuang Yu [33], which is 0.42% higher in OD, and 2.46% in the more difficult OC segmentation task. On REFUGE dataset, our method is also better than that of M-net [16] using this dataset.…”

“…In order to verify that our method is better than other methods, we compare our proposed method with the state-of-the-art approaches, such as pOSAL framework [27], GL-Net [28], M-Net [16], Stack-U-Net [29], WGAN [30], two-stage Mask R-CNN [31], multi-modal self-supervised pretraining network [32], Shuang Yu [33] and A. Sevastopolsky [10]. Additionally, we compare with the Fully convolutional network U-Net [15].…”