One of the biggest differences between medical imaging data and the rest of our everyday images is that a lot of them are in 3D, especially when dealing with DICOM series data. A DICOM image consists of many 2D slices that make up a scan or a specific part of the body.
So how do you build a deep learning solution for this kind of data? In this article, we will introduce 6 neural network architectures that can be used to train deep learning models on 3D medical data.
3 d U-Net
The U-Net architecture is a powerful medical image segmentation model. 3D U-Net extends the classic U-Net model to 3D segmentation. It consists of an encoding (downsampling) path and a decoding (upsampling) path.
The encoding path captures the context in the input image, while the decoding path allows precise localization. 3D U-Net is very effective in dealing with the 3D nature of volumetric images.
V-net
The V-Net architecture is another 3D convolutional neural network for volumetric image segmentation. Similar to U-Net, V-Net has an encoder-decoder architecture, but it uses full-resolution 3D convolutions, so it is more computationally expensive than U-Net.
HighResNet
It uses a series of 3D convolutional layers with residual connections. The model is trained end-to-end to process the entire 3D image at once.
EfficientNet3D
This is a 3D improvement on the EfficientNet architecture, which is not as commonly used for 3D segmentation as U-Net or V-Net, but it can be considered if computing resources are limited because of its good trade-off between computational cost and performance.
Attention U-Net
This is a variant of U-Net that incorporates an attention mechanism that allows the network to focus on certain parts of the image that are more relevant to the task at hand.
DeepMedic
This is a 3D CNN using a dual path, one with normal resolution and the other with downsampled input, which combines local and larger contextual information.
Summarize
In this article, we describe some of the deep learning models used by the medical imaging industry when processing 3D MRI and CT scans. These neural networks are designed to receive 3D data as input to learn the intricacies of specific body parts in DICOM series.
The following are all the model design papers in this article. If you are interested, you can directly view the papers for more detailed information.
https://avoid.overfit.cn/post/b874f819150d42fa8a64b967e102b99c
作者：Nour Islam Mokhtari