This Project includes 2 Task :
- Covid 19 Infection Segmentation
- Covid 19 Lungs Segmentation
Project pipeline:
1. As mentioned earlier, slices are first processed with CLAHE (Contrast Limiting Adaptive Histogram Equalization)
The impact of the filter on the sharpness of the image is clearly identifiable.
2. After CLAHE, I've cropped the ROI:
3. The next step is removing incomplete and fauty images (empty masks, e.g.)
- As a model I've used the baseline UNet.
- For training subsets I've used the simple transforms (Horizontal/Vertical flips) from Albumentations
- I've used Catalyst runner for train (it's very comfortable to use) with wandb logger
- 30 epochs + combination of losses (DiceLoss, IouLoss, BCELoss) + Adam optimizer + CosineAnnealingWarmRestarts scheduler for training loop
For baseline we have:
- Task 1 (Covid 19 Infection Segmentation) :
- IoU = 0.60
- DICE = 0.75
- Task 2 (Covid 19 Lungs Segmentation) :
- IoU = 0.71
- DICE = 0.83