brain-tumor-segmentation-0002#

Use Case and High-Level Description#

This model was created for participation in the Brain Tumor Segmentation Challenge (BraTS) 2019. It has the UNet architecture trained with residual blocks.

Specification#

Metric

Value

Type

Segmentation

GFLOPs

300.801

MParams

4.51

Source framework

PyTorch*

Accuracy#

See BRATS 2019 Leaderboard. The metrics for challenge validation (Dice_WT, Dice_TC, Dice_ET) differ from the metrics reported below (which are compartible with input labels):

  • WT (whole tumor) class combines all three tumor classes:

    • necrotic core / non-enhancing tumor

    • edema

    • enhancing tumor

  • TC (tumor core) combines the following classes:

    • necrotic core

    • non-enhancing tumor

  • ET (enhancing tumor)

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

Mean:

  • Dice index for “overall”: 91.5%

  • Dice index for “necrotic core / non-enhancing tumor”: 61.1%

  • Dice index for “edema”: 80.6%

  • Dice index for “enhancing tumor”: 79.4%

Median:

  • Dice index for “overall”: 92.7%

  • Dice index for “necrotic core / non-enhancing tumor”: 64.5%

  • Dice index for “edema”: 83.5%

  • Dice index for “enhancing tumor”: 86%

NOTE: The accuracy achieved with ONNX* model adapted for OpenVINO™ can slightly differ from the accuracy achieved with the original PyTorch* model since the upsampling operation was changed from the trilinear to nearest mode.

Input#

The model takes as an input four MRI modalities T1, T1ce, T2, Flair. Find additional information on the BraTS 2019 page and wiki. In the preprocessing pipeline, each modality should be z-score normalized separately. The input tensor is a concatenation of the four input modalities.

Original Model#

MR Image, name - 0, 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, T1ce, T2, Flair.

Converted Model#

MR Image, name - 0, 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, T1ce, T2, Flair.

Output#

Original Model#

Probabilities of the given voxel to be in the corresponding class, name - 304, shape - 1, 3, 128, 128, 128, output data format is B, C, D, H, W, where:

  • B - batch size

  • C - channel

  • D - depth

  • H - height

  • W - width

The channels are ordered as whole tumor, tumor core, and enhancing tumor.

Converted Model#

Probabilities of the given voxel to be in the corresponding class, name - 304, shape - 1, 3, 128, 128, 128, output data format is B, C, D, H, W, where:

  • B - batch size

  • C - channel

  • D - depth

  • H - height

  • W - width

The channels are ordered as whole tumor, tumor core, and enhancing tumor.

Download a Model and Convert it into OpenVINO™ IR Format#

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

An example of using the Model Downloader:

omz_downloader --name <model_name>

An example of using the Model Converter:

omz_converter --name <model_name>

Demo usage#

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: