Eformer It's a low dose CT Image reconstruction work , For the first time, it will Transformer Used in medical image denoising ,ICCV2021 Of paper, Worth reading .

Know that the account with the same name is released synchronously .

One 、 Architecture and contribution

and Uformer equally , take LeWin Transformer Joined the UNet in , The difference is that it will go through Sobel Convolution Processed images concat To UNet At different stages of encode and decode In the process . Residual learning is used , The residual of the original image is the result of prediction and denoising . Note that down sampling and up sampling use convolution and deconvolution （ Be careful to avoid the chessboard effect , I'll talk about it later ）.

Two 、 Key details

The main thing to understand is Sobel Convolution and LeWin Transformer, The latter I am Uformer note Has been recorded in , Simply put the formula ：

2.1,Sobel Convolution

Sobel Convolution The function of is to get edge-enhanced images, The effect is shown below ：

I didn't see the details , The references to them in the article are placed below ：

[19] Tengfei Liang, Yi Jin, Yidong Li, and Tao Wang. Edcnn: Edge enhancement-based densely connected network with compound loss for low-dose ct denoising. 2020 15th IEEE International Conference on Signal Processing (ICSP), Dec 2020.
[24] Irwin Sobel. An isotropic 3x3 image gradient operator. Presentation at Stanford A.I. Project 1968, 02 2014.

Sobel Convolution The following activation function is GeLU.

2.2, Down sampling and up sampling

use 3×3 The convolution of is down sampled ,stride by 2,padding by 1. The author said no pooling The reason is that details may be lost , So using strided conv Let's take a sample .

Deconvolution is used for up sampling , The author mentioned the chessboard effect ：

Chessboard effect , The source is the deconvolution process , When the convolution kernel size cannot be divided by the step size , Deconvolution will cause overlapping problem , When inserting zero , There will be some numerical effects on the output results , It's like a chessboard .

So the convolution kernel size should be divisible by the step size , Adopted by the author 4×4 Convolution kernel ,stride by 2.

2.3, Loss function

Using two loss functions , One is MSE, One is perceived loss .

MSE：

MSE Loss can cause excessive smoothing and image blur , So the author didn't just use it .

ResNet based MSP(Multi-scale Perceptual)：

$x_i-R(x_i)$ Is the prediction result of the network ,$y_i$ yes ground-truth,${\varphi }_s$ Express ResNet, Different s Represent different scales of ResNet, Altogether C Kind of s, So it is multi-scale.

The final loss function is as follows ：

The author used different methods to compare the loss ：

3、 ... and 、 experiment

stay Mayo do , First use Pydicom Remove pixels from Dicom File extraction to Numpy array in , The pixel value is then scale To 0-1 Between . Other data sets and training details are in paper The first 7 page .

Comparison with other methods ：

The effect picture put by the author ：