First, the project uses a neural network model to classify the brain MRI images into these four categories :
- Glioma
- Pituitary
- Meningioma
- No Tumor
Then I started working on the No Tumor dataset to see whether certain rare tumor cases were erringly classified as having no tumor. I used a Generative Adversarial Network (GAN) model to augment more images to mitigate bias as much as possible and then checked those images for certain characteristics that might signal the presence of a rare case of the tumor. In the end, a binary classification model (here CNN) was trained with the help of a labeled dataset ( No Tumor / Rare Case ) and then tested on an unseen chunk of the same dataset.
The GAN model has not been trained for more than one epoch (as of now) due to lack of certain resources. The code will be updated as soon as I run it on a better GPU.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.