1. Model improvement
1. Introduce FCN
2. Introduce U-Net
3. Introduce Res-UNet
2. An improved multi-scale U-Net segmentation method
a. Multi-scale target segmentation
integrates multi-scale input mechanism into U-Net network Middle
and downsampling module : The training images are made into four images according to the ratio of 8:4:2:1, and the one, two, three, and four down-sampling layer of U-Net are merged
. Up-sampling module : feature maps of each scale Make predictions, perform a kind of deep supervision on the prediction results (similar to the deep supervision in U-Net+++), transform the labels into 8:4:2:1 according to the ratio, and compare the predicted results of each layer with the corresponding labels Comparison
b. Attention mechanism
b1. Feature pyramid attention module
b2. Attention channel (channel attention mechanism)
The output result of each layer in the upsampling stage is introduced into the channel attention mechanism
b3. Spatial attention mechanism
In the up-sampling module, a spatial attention mechanism is introduced
. 2. The overall model. The
green module is the feature pyramid attention module
3. Loss function design
4. Comparative experiment
1. Data preprocessing: using adaptive median filtering Processing the original image
Data enhancement: rotation and horizontal flip
2. Optimizer Adam, lr=5e-4, bs=8, 100epoch
3. Experiment one (Dice loss function used):
①FCN ②Improved
FCN
4. Experiment two ( The Dice loss function used (that is, the ablation experiment):
①UNet ②Multi
-scale input mechanism + UNet ③Multi
-scale input mechanism + feature pyramid structure + UNet ④Multi
-scale input mechanism + feature pyramid structure + channel space attention mechanism + UNet
5. Experiment 3 (Dice loss function used) (compare the network model of this article with other network models):
①FCN
②UNet
③Res- UNet ④Multi-
scale input mechanism + feature pyramid structure + channel space attention mechanism + UNet
6. Experiment 4 (verification design Loss function) to
verify which value of α, β, γ is better, and compare it with the original Dice loss function
