Overview / Usage

This repo utilize a ensemble of 2-D and 3-D fully convoultional neural network (CNN) for segmentation of the brain tumor and its constituents from multi modal Magnetic Resonance Images (MRI). The dense connectivity pattern used in the segmentation network enables effective reuse of features with lesser number of network parameters. On the BraTS validation data, the segmentation network achieved a whole tumor, tumor core and active tumor dice of 0.89, 0.76, 0.76 respectively.

Methodology / Approach

• Our algorithm makes use of the ANTs framework for mask generation. First, call deepSeg class builds ANTs framework locally in ~/.DeepBrainSeg

• First, call deepSeg downloads all pre-trained models locally in ~/.DeepBrainSeg

• Final segmentation is the result of an ensemble of 4 different models:

  • ABLNet (modelABL.py, Air brain Lesion Network)
  • 3DBrainNet (model3DBNET.py, 3D multiresolution CNN)
  • Tiramisu2D (modelTis2D.py, 57 layered 2D CNN)
  • Tiramisu 3D (modelTir3D.py, 57 layered 3D CNN)

• Extensive documentation will be uploaded soon, along with transfer learning framework

• More details about network architecture and training procedure can be found here

Technologies Used

Pytorch, numpy, pandas, nibabel,simpleITK

Repository

https://github.com/koriavinash1/DeepBrainSeg