Brief Review — Analysis of Lung Scan Imaging Using Deep Multi-Task Learning Structure for Covid-19 Disease
U-Net With Classification Branch for Multi-Task Learning (MTL)
Analysis of Lung Scan Imaging Using Deep Multi-Task Learning Structure for Covid-19 Disease, U-Net+MTL, by Islamic Azad University,
2023 IET-IPR (Sik-Ho Tsang @ Medium)Biomedical Image Multi-Task Learning
2018 … 2020 [BUSI] [Song JBHI’20] [cGAN JESWA’20] 2021 [Ciga JMEDIA’21] [CMSVNetIter] 2023 [FFANet+MTL]
==== My Other Paper Readings Are Also Over Here ====
- A multi-task U-Net model is proposed which can perform automatic classification-segmentation for screening COVID-19 pneumonia by using chest CT imaging.
- This model includes a common encoder for feature representation, one decoder for segmentation, and a multi-layer perceptron for classification, respectively.
Outline
- Proposed Model Architecture
- Results
1. Proposed Model Architecture
1.1. Model Architecture
- At first, images are converted into features so that they can be entered into the encoder, and form the latent space.
- The variables in latent space are entered into the decoder for segmentation task and fully connected layers for classification task.
- U-Net is used where the encoder increases the number of filter from 64 to 1024 using 3×3 conv.
- The convolution block followed by skip connection and using stride of 2 is used instead of pooling operation.
- The decoder works on upsampling to recover the segmentation mask with the same size of input.
- The multilayer perceptron consists of two dense layers with 128 and 64 neurons, respectively, with a Dropout of 0.5 and the activation function ReLU.
- Both outputs use Sigmoid activations.
1.2. Loss Function
- Cross entropy loss is used for classification:
- Dice loss is used for segmentation:
- Combined loss is used for final loss.
2. Results
2.1. Comparisons with U-Net
The proposed multi-task model had the best performance among other previous simple U-Net methods.
2.2. Classification
- Different new models, as above, were compared for classification.
The proposed model could obtain the best outcomes in all of the evaluation metrics.
2.3. Segmentation
The model had the best result compared to those in the previous studies.
The proposed model was able to acceptably distinguish normal and infected areas. Also, the model was able to separate the areas involved in other infections from the areas involved in the COVID-19 infection.
