kaggle link -> https://www.kaggle.com/code/banddaniel/retina-vessel-segmentation-w-tpu-test-dice-0-75

I have used the following methods.

• I used two image processing methods for images (Green Channel Conversion[2] , Histogram Equalization[3])
• I used a morphological image processing method for masks (Dilation[4])
• Dice coefficient[5] implementation,
• The project took place using Google TPU,
• Custom layers for encoding and decoding,
• Custom callback class that used predicting a sample from the test dataset during training
• 1000 epochs for training (of course, although this number is very high, the metrics(dice, loss) continued improvement during training)

1. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation (Version 1). arXiv. https://doi.org/10.48550/ARXIV.1505.04597
2. Rathod, Deepali & Manza, Ramesh & Rajput, Yogesh & Patwari, Manjiri & Saswade, Manoj & Deshpande, Neha. (2014). Localization of Optic Disc and Macula using Multilevel 2-D Wavelet Decomposition Based on Haar Wavelet Transform. International Journal of Engineering Research & Technology (IJERT)
3. https://en.wikipedia.org/wiki/Histogram_equalization
4. https://homepages.inf.ed.ac.uk/rbf/HIPR2/dilate.htm
5. https://en.wikipedia.org/wiki/Sørensen–Dice_coefficient