研究目的
To introduce a new convolutional neural network structure for medical image denoising based on wavelet domain to overcome limitations in computational efficiency and manually chosen parameters in existing methods.
研究成果
The DWDN model effectively denoises medical images by combining wavelet transform with deep learning, achieving higher PSNR values than existing methods like BM3D, WNNM, and DnCNN. It excels in preserving image details and simplifies learning through residual and cascade training. Future work could explore applications to other noise types and larger datasets.
研究不足
The dataset is relatively small with only 742 images, which may limit generalization. The method is specifically tested with Gaussian noise and may not perform as well with other noise types. Computational efficiency and parameter tuning in deeper networks could be challenging without the cascade training approach.
1：Experimental Design and Method Selection:
The study employs a deep wavelet denoising net (DWDN) that integrates wavelet transform with convolutional neural networks for adaptive threshold selection and noise reduction. It uses Haar wavelet transform to process images in the frequency domain while retaining spatial information, and incorporates residual blocks (ResBlocks) to mitigate gradient issues in deep networks. A cascade training algorithm is used to gradually deepen the network.
2：Sample Selection and Data Sources:
742 medical stomach pathological images of size 2048x2048 pixels are used. These are cropped into 256x256 sub-images with 128-pixel overlaps to increase the dataset size.
3：List of Experimental Equipment and Materials:
No specific equipment or materials are mentioned; the focus is on computational methods using neural networks and image processing techniques.
4：Experimental Procedures and Operational Workflow:
Steps include: adding Gaussian noise to original images to create noisy images (NI), applying Haar wavelet transform to NI and pure images (PI) to generate input-label pairs, training the DWDN model with these pairs using residual learning and cascade training, and applying inverse wavelet transform to the network output to obtain denoised images.
5：Data Analysis Methods:
Performance is evaluated using Peak Signal-to-Noise Ratio (PSNR) to measure similarity between denoised and original images. Comparisons are made with BM3D, WNNM, and DnCNN methods.
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
