Instance segmentation of histology images from human kidney tissue slides
The data is from taken from a kaggle competition, the goal is to segment instances of microvascular structures on a histology images.
Project starts with Data preprocess/analysis notebook, in wich methods for data reconstruction, analysis and image augmentation are presented.
For main model I used UNet architecture with pre-trained Mix Vision Transformer (backbone from SegFormer) as an encoder. In the second notebook - 'WSI reconstruction' you can find model training and CV score estimation with K-fold CV method.
I tried different encoder models like ResNet34, ResNet101, VGG16, DenseNet etc., the best results I got with Vision Transformer model.
For the last model evaluation results, check the Model evaluation notebook.
Initial CV scoring got me target class IoU=0.33.
After hyperparameter tuning, using L2 and Dropout regularization, I got mean score around 0.4.
To separate every instance of target class, I used the Connected components method.