Hi,

For MS level study, I think it's important to have plenty of resources that can get a student started fairly quickly. This seems to be the case for brain tumor analysis. Some helpful factors:

(1) Dataset availability: The BRATS datasets http://braintumorsegmentation.org/ have evolved over the course of several years. There have been a variety of neural network and other methods applied to these datasets, meaning a prospective student has a good number of baselines against which to compare their system.

(2) For deep convolutional networks, there is both existing research available and even some code availability. I'll mention some instances below.

(3) Most of the major deep learning networks now support the key covnet operators when applied to 3D data, so experimenting with such systems is simpler than in the past.

More concretely, I see there has been some work on brain tumor segmentation for the 2015 BRATS dataset using an architecture one of my masters students has been experimenting with successfully. This is the U-net architecture, originally proposed for segmentation of electron microscope images of neuronal structures (the ISBI challenge dataset). There appears to be code related to the BRATS dataset paper available at https://github.com/zsdonghao/u-net-brain-tumor . For reference, I'm attaching the original U-net paper and the recent paper of U-net applied to the BRATS dataset. The BRATS paper just mentioned attempts the 2D segmentation problem, but there is also code available for U-net in 3D operating on BRATS at another github repository https://github.com/ellisdg/3DUnetCNN  (seemingly no associated paper with this repository, but other readers should chime in if they've seen one).

U-net and fully convolutional neural networks like it in general are relatively simple to understand, but have rich and interesting behaviour. They are also open to being adapted in novel ways. That makes them a good platform for masters work, but also for continuing into higher degrees (possibly branching into other deep learning methods / applications).

I should maybe emphasise that there are other approaches to the 2D/3D segmentation problem using deep neural networks that have been employed, for example, on BRATS, but I thought I'd concentrate on a variant I have supervised a project on.

Feasibility for masters study probably depends on how long the period of study is at a particular university. Is it 1, 2 or 3 years? How much of that is project work (partially or entirely)? My students are on 2 year masters degree programs, with most of that being project research activities. That leaves some room for exploration (I'm attaching the conference papers of the student I mentioned earlier, Stiaan Wiehman, as a point of reference).

If the masters degree project time is shorter (a year?), I think the important thing is to define the scope of the project clearly and narrowly. Perhaps this could involve implementing and testing a novel preagreed upon variant on an existing architecture, and depending on the result elaborate if there is time? It seems doable, as long as the student has a good/relevant background coming into the masters program.

Conference Paper U-Net: Convolutional Networks for Biomedical Image Segmentation

Conference Paper Semantic Segmentation of Bioimages Using Convolutional Neural Networks

Conference Paper Unsupervised pre-training for fully convolutional neural networks

Conference Paper Automatic Brain Tumor Detection and Segmentation Using U-Net...

i need a matlab code for segmentation of brain tumors in MRI images using BRATS dataset @ Hendrik De Villiers

Hi

any one help me to doxnload the dataset Isles 2015 or 2016 or 2017 or Atlas, for the segmentation of ischemia stroke brain?

[email protected]