UWA-Image Segmentation

Kaggle medical image segmentation project. End-to-end from training to inference. Tensorflow/keras implementation

---

Dataset found on: https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/data

---

Ensure that PATH_DIR points to the train/ folder in the dataset given above.

2.5D UNet with processed skip connections

Multiple models investigated + various architectural blocks (commented out in jupyter notebook): e.g:

• https://arxiv.org/abs/1909.00166
• https://arxiv.org/abs/1902.04049
• https://arxiv.org/abs/2003.05056

Augmentations:

• Colour normalization via CLAHE
• Colour augmentation ("possibly omit in the context of 2.5D inputs)
• Scale/zoom/rotate augmentation
• Flip augmentation
• Coarse cutout (shamelessly stolen from another person on kaggle)