Brief Review — IBA-U-Net: Attentive BConvLSTM U-Net with Redesigned Inception for medical image segmentation
IBA-U-Net, Uses Dense Block, & LSTM with Attention in U-Net
IBA-U-Net: Attentive BConvLSTM U-Net with Redesigned Inception for medical image segmentation,
IBA-U-Net, by Nanchang University, and Carleton University, 2021 J. Comput. Biology and Medicine, Over 10 Citations (Sik-Ho Tsang @ Medium)
Medical Imaging, Medical Image Analysis, Image Segmentation, U-Net
4.2. Biomedical Image Segmentation
2015–2021 [Expanded U-Net] [3-D RU-Net] [nnU-Net] [TransUNet] [CoTr] [TransBTS] [Swin-Unet] [Swin UNETR] [RCU-Net] 2022 [UNETR]
My Other Previous Paper Readings Also Over Here
- The encoder-decoder Attentive BConvLSTM U-Net with Redesigned Inception (IBA-U-Net) is proposed.
- Integrating the BConvLSTM block and the Attention block to reduce the semantic gap between the encoder and decoder feature maps to make the two feature maps more homogeneous.
- Factorizing convolutions with a large filter size by Redesigned Inception, which uses a multiscale feature fusion method to significantly increase the effective receptive field.
1. Attentive BConvLSTM U-Net with Redesigned Inception (IBA-U-Net)
- Medical image segmentation, using encoder-decoder architecture, such as U-Net, combine local information based on high-resolution images and low-resolution global features to distinguish foreground and background, as above.
- Redesigned Inception (RI) block is proposed to enhance the features throughout the network.
- (It is assumed that we know inception block, dense block, LSTM, and attention operation below.)
1.1. Redesigned Inception (RI) Block
- In original Inception block, independent convolutional paths are used to extract features with different receptive fields.
- The proposed RI block uses Dense block concept as in DenseNet to extract the multiscale details of different scales, and it is used to replace the traditional two 3×3 convolutional layers in U-Net.
1.2. Attentive BConvLSTM Block
- Instead of the simple skip connection in U-Net, the proposed Attentive BConvLSTM block, integrating the BConvLstm block and the Attention block, is used to reduce the semantic gap between the encoder and decoder feature maps.
- First, the BConvLSTM block provides hidden state tensors for the Attention block.
- Then, the Attention block pays different degrees of attention to the hidden state tensors output by BConvLSTM to highlight the salient features in skip connections.
2.1. Lung Segementation
The IBA-U-Net achieves 99.55% in Sen, 99.73% in Acc and 99.45% in Auc, outperforms the state-of-the-art methods on the lung segmentation dataset using different performance metrics.
- Adding the RI block and BA block to the original U-Net greatly increased the segmentation accuracy.
- U-Net + RI exceeds U-Net + Inception in all four indicators, which shows that the proposed RI block is more suitable for medical image segmentation.
- The U-Net + BA block also exceeds only using U-Net + Attention or U-Net + BConvLSTM.
The proposed IBA-U-Net network showed the best performance, among which the F1-Score and precision significantly improved.
U-Net has many incorrect segmentations (areas indicated by red ovals), whereas IBA-U-Net segments the images reliably without making false predictions.
2.4. Skin Disease Segmentation
IBA-U-Net achieves 82.91% in Sen, 94.40% in Acc and 90.13% in Auc, which are better than other methods on the skin disease dataset.
The proposed network surpasses U-Net in all four indicators, proving that the network proposed in this paper has a good performance on challenging.
- When the background and the segmented image are similar in color, U-Net makes some incorrect predictions. The rougher the background, the more incorrect predictions U-Net makes.
At the same time, U-Net + BA and IBA-U-Net show finer segmentation.
2.5. Retinal Vessel Segmentation
IBA-U-Net achieves 78.58% in Sen, 95.50% in Acc and 97.78% in AUC, which outperforms the other methods.
The segmentation indicators of IBA-U-Net and U-Net are not much different.
- There are some small background pixels in the ground truth. The U-Net + BA network and IBA-U-Net network successfully segmented part of the background pixels, but U-Net often missed them.
In addition, the results of IBA-U-Net segmentation show that the network is more immune to noise in the image.
The addition of the RI block can greatly reduce the network parameters while increasing the F1-Score and intersection over union.