DEFN: Dual-Encoder Fourier Group Harmonics Network for Three-Dimensional Macular Hole Reconstruction with Stochastic Retinal Defect Augmentation and Dynamic Weight Composition
DEFN: Dual-Encoder Fourier Group Harmonics Network for Three-Dimensional Macular Hole Reconstruction with Stochastic Retinal Defect Augmentation and Dynamic Weight Composition.
Xingru Huang, Yihao Guo, Jian Huang, Zhi Li, Tianyun Zhang, Kunyan Cai, Gaopeng Huang, Wenhao Chen, Zhaoyang Xu, Liangqiong Qu, Ji Hu, Tinyu Wang, Shaowei Jiang, Chenggang Yan, Yaoqi Sun, Xin Ye, Yaqi Wang
Hangzhou Dianzi University
Figure 1: The System structure.
Figure 2: The network structure of DEFN.
We proposed DEFN, a 3D OCT segmentation network for eye diseases with unobvious characteristics and easy to be interfered with, such as macular hole and macular edema.
This repository contains the official Pytorch implementation for DEFN networking and DWC Loss, as well as the pre-trained model for DEFN.
Figure 3: The FuGH Module.
The Fourier Group Harmonics (FuGH) module enhances noise reduction in OCT images by employing Fourier transformation for feature extraction in the frequency domain, enabling targeted noise filtration and efficient processing of periodic patterns with reduced computational complexity.
Figure 4: The S3DSA Module.
The Simplified 3D Spatial Attention (S3DSA) module improves the segmentation of macular holes and edema in retinal images by optimizing spatial attention mechanisms. It refines focus on crucial regions, enhancing segmentation quality and computational efficiency.
Figure 5: The HSE Module.
The Harmonic Squeeze-and-Excitation Module (HSE) combines Fourier Group Harmonics (FuGH) and Squeeze-and-Excitation (SE) blocks to enhance the segmentation of macular holes and edema, by extending the model's view field and recalibrating feature weights.
We have provided the GitHub links to the PyTorch implementation code for all networks compared in the experiments herein.
3D UX-Net, nnFormer, 3D U-Net, SegResNet, SwinUNETR, TransBTS, UNETR, DeepResUNet, ResUNet, HighRes3DNet, MultiResUNet, SegCaps, V-Net
Figure 6: The Training Results using Isolated Strategy.
Segmentation results employing the isolated macular hole injection method, comparing the proposed DEFN, DEFN+DWC Loss, and prior segmentation models. The evaluation spans four classes: All (Average across all classes), MH (Macular Hole), ME (Macular Edema) and RA (Retina). The best values for each metric are highlighted in red, while the second-best values are highlighted in blue, and the values of our model are bolded.
Figure 7: The Training Results using Comprehensive Strategy.
Segmentation results employing the comprehensive macular hole injection method, comparing the proposed DEFN, DEFN+DWC Loss, and prior segmentation models. The evaluation spans four classes: All (Average across all classes), MH (Macular Hole), ME (Macular Edema) and RA (Retina). The best values for each metric are highlighted in red, while the second-best values are highlighted in blue, and the values of our model are bolded.
Figure 8: The Fine-tuning Results using Isolated Strategy.
Segmentation results of fine-tuning after isolated macular hole injection training, comparing the proposed DEFN, DEFN+DWC Loss, and prior segmentation models. The evaluation spans four classes: All (Average across all classes), MH (Macular Hole), ME (Macular Edema), and RA (Retina). The best values for each metric are highlighted in red, while the second-best values are highlighted in blue, and the values of our model are bolded.
Figure 9: The Fine-tuning Results using Comprehensive Strategy.
Segmentation results of fine-tuning after comprehensive macular hole injection training, comparing the proposed DEFN, DEFN+DWC Loss, and prior segmentation models. The evaluation spans four classes: All (Average across all classes), MH (Macular Hole), ME (Macular Edema), and RA (Retina). The best values for each metric are highlighted in red, while the second-best values are highlighted in blue, and the values of our model are bolded.
Figure 10: The Ablation Results.
The ablation study examined the backbone and modules of DEFN and our methods, including DEFN (DEFN backbone), HSE (Harmonic Squeeze-and-Excitation Module), FuGH (Fourier Group Harmonics), IMHI (Isolated Macular Hole Injection), CMHI (Comprehensive Macular Hole Injection), and DWC (DynamicWeightCompose). Optimal metric values are highlighted in red, while the next best are in blue.
Figure 11: The 3D Reconstruction Results.
The display includes five cases, each with original and reconstructed images. The first row shows pre-segmentation images, rows two to five show reconstructions from five grid arrangements, and the sixth row shows a top-view of the reconstructions. Yellow indicates macular holes, and green signifies macular edema.