Using Deep Learning to segment a tumor in a Brain MR image.

Authors:

1. Duong Le Tuong Khang
2. La Truong Hai
3. Nguyen Huu Khang

This github includes preprocessing algorithms, segmentation models and post-processing method

Dataset:

We use dataset from a paper: Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm

Preprocessing

We uses this github to remove skull in MRI image to increase accuracy.
We implement CLAHE to image and resize it to common frame (256x256).

Segmentation Models

We compare some models like Unet, Unet++ and analysis advantages and drawbacks between them. Furthermore, we distinguish a number of loss functions: BCE-Dice loss, tversky loss, focal tversky loss and Jaccard Loss. Also, in this project, we study about how adam and sgd optimization affected model and provide a summary about them.

Post-processing

When we had a model, we recognized some drawbacks. Initially, some holes and small regions emerged in masks. To solve this problem, we apply a algorithms that it finds a connected components and deletes follow to a threshold.

Depoy web

Source code in Web deployment folder.