U_Net for medical image segmentation

In this repo I have implemented the U_Net architecture as proposed in this paper - U-Net: Convolutional Networks for Biomedical Image Segmentation . The architecture in the paper is as follows: I have not implemented the architecture as it is but made some minor changes like adding dropouts etc. to improve my accuracy. The final architecture I used is as follows:

arch.py - file contains the code for the architecture

data.py - file contains all the neccesary code for data preprocessing, one can make changes in this file according to their own dataset

To train-

python main.py

Requirement:

Tensorflow>=1.13.1

Dataset used here:

2018 Data Science Bowl | Kaggle

I have trained this architecture for cell nucleus detection from images.Masks of the cell nuclei are provided with the dataset to train our model for segmentation purpose.Training image example-

Mask of the above image-

Following are the test results: