- This Project is created by the joint efforts of Rohit Sharma , Abdul Mugeesh and Kanishk Nama . Kanishk Nama is responsible for preparing and loading the dataset, Rohit Sharma is responsible for developing the model while Abdul Mugeesh is responsible for calculating and loading the results. Rest of the work is done together.
DRIVE - Digital Retinal Images for Vessel Extraction
We have done segmentation of blood vessels from their respective retinal images. As this is a segmentation problem, we have used U-net architecture for the segmentation purpose. This is a binary classification task as at each pixel, we have to classify whether that pixel is feature or not.
We have used DRIVE dataset for training and testing purpose, this dataset consists of 40 images, 20 for training purpose and 20 for testing.
- We are using DenseBlock U-net, the architecture is same as U-net but only difference is we are replacing 1 convolutional layer with Dense layer.
base : store template class of data_loader trainer infer model
model : store the definition of semantic segmention model
infer : how to predict
trainer : the way to train your model
- First Clone the github repository.
- For Training Purposes First, you have to set all the paths in the files wherever necessary, in config directory and then run main_train.py
- For Testing Purpose You have to run the main_test.py file, All the results will be stored in experiments/VesselNet/checkpoint folder.
- To modify the model, you can first change the path in all the python files and then change segmentionmodel.py or to add the dense layer change thr denseunet.py file.
We have written a Research Paper based on our work and as a future part we are aspiring to add more Dense layers and performing hyper-parameter tuning.