Deep convolutional neural network implementation for brain tumor segmentation.
In this project I pre-processed MRI scans by applying skull-stripping techniques and eventually built an extremely lightweight convolutional neural network with Keras for brain tumor segmentation.
The dataset used was created by Buda et al. (https://arxiv.org/abs/1906.03720). The authors used a U-Net model for brain tumor segmentation. U-Net has 7m paramenters. Our goal was to develop a lightweight model that would yield similar performance of the author's U-Net implementation (82% Dice score).
Our model, a modified MobileNetv3, achieved 85% Dice score, outperforming the original U-Net implementation of the authors whilst having 0.05% of U-Net's paramenters (350k v 7m params).
Project report: ISPR_PROJECT_REPORT
Project presentation: ISPR_PROJECT_SLIDES