研究目的
To develop a multi-input fully-convolutional network (MIFCN) for denoising optical coherence tomography (OCT) images by exploiting correlations among nearby OCT images to reduce noise while preserving textures and layer structures.
研究成果
The proposed MIFCN method effectively reduces noise in OCT images by leveraging correlations among nearby images through a neural network architecture with a weighted averaging module. It outperforms state-of-the-art denoising methods in both quantitative metrics and visual quality, preserving layer structures and textures with fewer artifacts. Future work could integrate segmentation information and extend the method to other applications like image interpolation.
研究不足
The method relies on the availability of multiple nearby OCT images for effective denoising; performance may degrade if such correlations are weak. The parameter h in the weighting mechanism requires empirical tuning, which could be dataset-dependent. Training data size is limited, potentially leading to overfitting with more complex architectures. The approach is specific to OCT images and may not generalize well to other medical imaging modalities without adaptation.
1：Experimental Design and Method Selection:
The study uses a multi-input fully-convolutional network (MIFCN) architecture with multiple branches, each processing a noisy OCT image, followed by a weighted averaging module inspired by nonlocal mean weighting to fuse outputs. The network is trained end-to-end using a loss function that ensures consistency between outputs and contributions from each branch.
2：Sample Selection and Data Sources:
The dataset consists of spectral domain OCT (SDOCT) images from 28 subjects with normal and age-related macular degeneration eyes, publicly available from prior studies. Training uses 10 pairs of noisy and high SNR images; test set includes additional images with nearby OCT images provided.
3：List of Experimental Equipment and Materials:
A desktop PC with Intel i7-7700K CPU, 16 GB RAM, NVIDIA GeForce GTX 1080 Ti GPU; Bioptigen SDOCT imaging system for data acquisition.
4：Experimental Procedures and Operational Workflow:
Patches of size 15x15 pixels are extracted from high SNR images, and similar patches are found via nonlocal searching to create a training set. The network is trained for 60 epochs with data augmentation (flipping and rotation), using Adam optimizer with specified learning rates. During inference, the MIFCN processes multiple input images to produce denoised outputs.
5：Data Analysis Methods:
Performance is evaluated using quantitative metrics: PSNR, MSR, CNR, ENL, and Wilcoxon signed-rank test for statistical significance. Visual inspection is also used for qualitative assessment.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
