The goal of this project was to preprocess the given data and train a Convolutional Neural Network (CNN) model with Unet architecture to detect the coating layer on new images.
- Non-existing Dataset: The original dataset did not exist in a usable format, so we had to create one:
- Masks were generated by plotting polygons based on ROI lines.
- A custom K-Means algorithm was developed to assist in labeling.
- The process was refined through collaboration with scientists to ensure accuracy.
- using U-net architecture
- Augmentation: Various data augmentation techniques were employed.
- Data Splitting: The dataset was split into training, validation, and test.
- Optimization: Hyperparameters were optimized using Optuna.
- Shell Commands: Integration of the model into Fiji (ImageJ).
Left: Original image, Right: Image with predicted coating layer
- Fiji (ImageJ): https://imagej.net/software/fiji/
- U-Net: Ronneberger et al., "U-Net: Convolutional Networks for Biomedical Image Segmentation" – Paper
- Optuna for hyperparameter tuning: https://optuna.org
- Albumentations for augmentation: https://albumentations.ai
- segmentation_models.pytorch – A PyTorch library with segmentation architectures - U-Net