brain-tumor-segmentation-0001

Use Case and High-Level Description

This model was created for participation in the Brain Tumor Segmentation Challenge (BraTS) 2018. The model is based on the corresponding paper, where authors present deep cascaded approach for automatic brain tumor segmentation. The paper describes modifications to 3D UNet architecture and specific augmentation strategy to efficiently handle multimodal MRI input. Besides this, the approach to enhance segmentation quality with context obtained from models of the same topology operating on downscaled data is introduced. Each input modality has its own encoder which are later fused together to produce single output segmentation.

Specification

Metric Value
Type Segmentation
GFLOPs 409.996
MParams 38.192
Source framework MXNet*

Accuracy

The following accuracy metrics are measured on a brain tumor training subset of the Medical Decathlon dataset.

Mean:

  • Dice index for "overall": 92.4003%
  • Dice index for "necrotic core / non-enhancing tumor": 71.467%
  • Dice index for "edema": 82.0533%
  • Dice index for "enhancing tumor": 72.7001%

Median:

  • Dice index for "overall": 93.1653%
  • Dice index for "necrotic core / non-enhancing tumor": 77.1611%
  • Dice index for "edema": 85.3434%
  • Dice index for "enhancing tumor": 84.5571%

See the original repository.

Input

The model takes as an input four MRI modalities T1, T2, T1ce, Flair. The inputs are cropped, resamped and z-score normalized. You can find additional information on the BraTS 2018 page and wiki. In the preprocessing pipeline, all non-zero voxels are cropped and resampled to 128, 128, 128 resolution first. Then, each modality is z-score normalized separately. The input tensor is a concatenation of the four input modalities.

Original model

MR Image, name - data_crop, shape - 1, 4, 128, 128, 128, format is B, C, D, H, W, where:

  • B - batch size
  • C - channel
  • D - depth
  • H - height
  • W - width

The channels are ordered as T1, T2, T1ce, Flair.

Converted model

MR Image, name - data_crop, shape - 1, 4, 128, 128, 128, format is B, C, D, H, W, where:

  • B - batch size
  • C - channel
  • D - depth
  • H - height
  • W - width

The channels are ordered as T1, T2, T1ce, Flair.

Output

Original model

Probabilities of the given voxel to be in the corresponding class, name - softmax_lbl3, shape - 1, 4, 128, 128, 128, output data format is B, C, D, H, W, where:

  • B - batch size
  • C - channel
  • D - depth
  • H - height
  • W - width

With the following channels: background, necrotic core, edema and enhancing tumor.

Converted model

Probabilities of the given voxel to be in the corresponding class, name - softmax_lbl3, shape - 1, 4, 128, 128, 128, output data format is B, C, D, H, W, where:

  • B - batch size
  • C - channel
  • D - depth
  • H - height
  • W - width

With the following channels: background, necrotic core, edema and enhancing tumor.

Download a Model and Convert it into Inference Engine Format

You can download models and if necessary convert them into Inference Engine format using the Model Downloader and other automation tools as shown in the examples below.

An example of using the Model Downloader:

python3 <omz_dir>/tools/downloader/downloader.py --name <model_name>

An example of using the Model Converter:

python3 <omz_dir>/tools/downloader/converter.py --name <model_name>

Legal Information

The original model is distributed under the Apache License, Version 2.0. A copy of the license is provided in APACHE-2.0.txt.