Following parts,
- Data info
- Network architecture
- Loss function
- Other info
- Input: 2D image with size 96X96
- Labels (ground-truth): 3D coordinates. 3D coordinates will be put into 3D grid and the value of that entry equals 1 (show existence) or flux value (also consider flux information)
- cnn_residual
- cnn (initial cnn)
- Initial cnn: DeepSTORM3D + dropout
- cnn_no_dialte (0721): cnn without dilated conv layers, remove dilation rate
[2,4,8,16] - cnn_ReLU (0808):
LeakyReLU->ReLU - cnn_concatIM (0808): remove features
torch.cat((out, im),1), which concatenate output of layers with original input image, used in DeepSTORM3D, - cnn_residual (0808): the difference with concatIM as below,
out = layer(out) + out-> residual conv layer- deconv1 and deconv2 layers with
+outor not
- cnn_duc (0808): more than 1 version
- Loc3dResCNN: interpolate -> duc with aspp
- ResLocalizationCNN_DUC (0809): interpolate -> duc with plain conv, without leakyReLU and BN
- ResLocalizationCNN_DUC_v2 (0809): interpolate -> last layer duc
- cnn_hdc (0810): more than 1 version
- ResLocalizationCNN_HDC: dilation
[1,1,2,4,8,16] -> [1,1,2,5,9,17] - ResLocalizationCNN_HDC_v2 (v2): dilation changes
[1,1,2,4,8,16] -> 1,[1,2,5,9,17]*2
- ResLocalizationCNN_HDC: dilation
- MSE3D: Loss between output and ground-truth tensor. Entry values represent confidence or flux value.
$MSE3D = ||G\otimes\hat y - G\otimes\hat y||^2$ - Dice loss: Dice loss only cares about the existence.
$Dice=\frac{2|y\cap\hat y|}{|y|+|\hat y|}$ - Forward loss (0817): New fidelity term evaluate difference on 2D image,
$Forward=||A\otimes\hat y-I_0||^2$ .$A$ is the discrete 3D PSF matrix,$I_0$ is 2d observed image. - Implementation of forward loss (0830): Implement forward loss (new fidelity term) by fft, details can be found in GitHub and ZhiHu.
- Implementation of forward loss (0930): Implement fft with Neumann boundary condition, refered to paper.