Current baselines in biomedical image segmentation utilise fully-convolutional structures for the benefits of end-to-end trainability, size-invariance and efficiency. One such method is U-Net [1], a two-track contraction-expansion model which fuses features at different hierarchies with the objective of generating deep localisable features. Here, I introduce Corona-Net, a 3-part contribution dedicated to the classification, binary segmentation and multi-class segmentation of COVID-19. I first leverage the EfficientNet model [2] for COVID-19 diagnosis, achieving an accuracy of 94.00%. I then utilise and refine the U-Net architecture for both binary and 3-class (ground-glass, consolidation, pleural effusion) segmentation of COVID-19 symptoms, through inference on the 100-slice COVID-19 CT segmentation (chest axial CT) dataset 1, and 630-slice COVID-19 CT segmentation dataset 2 [3]. Through strong data augmentation and rigorous experimentation, I overcome the small dataset size (630) to achieve a Dice Loss of 79.65% (dataset 2) and 61.60% (dataset 1). Through Corona-Net, I aim to develop a reliable, visual-semantically balanced method for automatic COVID-19 diagnosis confirmation, in order to contribute to the fight against this pandemic.
- Classification
| Model | 1- BCE Loss | Optimiser | Learning Rate |
|---|---|---|---|
| Efficient-Net-b0 | 0.9251 | Adam | 1e-05 |
| Efficient-Net-b1 | 0.9400 | Adam | 1e-05 |
| Efficient-Net-b2 | 0.9096 | Adam | 1e-05 |
- Binary Segmentation
| Dice Coefficient | Rand Loss | Optimiser | Learning Rate |
|---|---|---|---|
| 0.5641 | 0.2167 | Adam | 1e-02 |
| 0.7374 | 0.1031 | Adam | 1e-03 |
| 0.7965 | 0.0766 | Adam | 1e-04 |
| 0.4745 | 0.1591 | Adam | 1e-05 |
- Multi-Class Segmentation
| Dice Coefficient | Rand Loss | Optimiser | Learning Rate |
|---|---|---|---|
| 0.5160 | 0.2490 | Adam | 1e-02 |
| 0.5900 | 0.2114 | Adam | 1e-03 |
| 0.6160 | 0.1985 | Adam | 1e-04 |
| 0.5001 | 0.2565 | Adam | 1e-05 |
[1] O. Ronneberger, P. Fischer, and T. Brox. U-Net:Convolutional Networks for Biomedical ImageSegmentation. INMICCAI, 2015.
[2] M. Tan and Q. V. Le. EfficientNet: Rethink-ing Model Scaling for Convolutional Neural Net-works In ICML, 2019.
[3] Medical Segmentation.com COVID-19 CT seg-mentation dataset, 2020.
[4] M. Buda, A. Saha, and M. A. Mazurowski. Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm. In Computers in Biology and Medicine, 2019.