nnUNet Lesion Segmentation

This repository contains the nnUNet model weights for lesion segmentation. The model was trained on large dataset with 3000+ cases, acquired on various scanners and field strengths. Lesion labels were acquired using semi-automatic lesion segmentation procedure.

Dataset

Dataset used for training consisted of FLAIR and T1 MRI images. Lesion masks were acquired semi-automatic.

Preprocessing

Preprocessing consisted of:

1. T1, FLAIR, SEGMENTATION registration to MNI ICBM 152 atlas
2. BACKGROUND removal
3. CROPPING to shape divisible by two: (192, 224, 192)

nnUNet Configs

Dataset Configuration

• Channel Names:
  • 0: FLAIR
  • 1: T1
• Labels:
  • 0: Background
  • 1: Lesion

Training Configuration

• Plans Name: nnUNetPlans
• Original Median Spacing (after transpose): [1.0, 1.0, 1.0]
• Original Median Shape (after transpose): [157, 183, 147]

3D Full Resolution Configuration

• Data Identifier: nnUNetPlans_3d_fullres
• Preprocessor Name: DefaultPreprocessor
• Number of Epochs: 1000
• Batch Size: 6
• Patch Size: [128, 128, 128]
• Median Image Size in Voxels: [157.0, 183.0, 147.0]
• Spacing: [1.0, 1.0, 1.0]
• Normalization Schemes: ZScoreNormalization (both channels)
• Use Mask for Normalization: True (both channels)
• UNet Class Name: PlainConvUNet
• UNet Base Number of Features: 32
• Number of Convolutions per Stage (Encoder): [2, 2, 2, 2, 2, 2]
• Number of Convolutions per Stage (Decoder): [2, 2, 2, 2, 2]
• Number of Pooling Operations per Axis: [5, 5, 5]
• Pooling Operation Kernel Sizes: [[1, 1, 1], [2, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2]]
• Convolution Kernel Sizes: [[3, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3]]
• UNet Max Number of Features: 320
• Resampling Function Data: resample_data_or_seg_to_shape
• Batch Dice: False

Usage

Before using this model you need nnUNet V2. For this, check original repository documentation on instalation https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/installation_instructions.md

Predicting

For predicting dataset your call command should look like:

INPUT_FOLDER=<path-to-input-folder-in-nnunet-compatible-format>
OUTPUT_FOLDER=<path-to-output-folder>

nnUNetv2_predict -d 100 -i $INPUT_FOLDER -o $OUTPUT_FOLDER -f 0 -c 3d_fullres

Fine-Tuning

For fine-tuning custom model your command should look like:

nnUNetv2_train <task-id> 3d_fullres <fold-id> -pretrained_weights "/<path-to-source>/Dataset100_FLAIR-T1/nnUNetTrainer__nnUNetPlans__3d_fullres/fold_0/checkpoint_best.pth"