Brats20 Segmentation Eval

brats20-segmentation-eval

Transfer Learning for Brain Tumor Segmentation — Wacker et al. (2019) (arXiv:1912.12452, 2019)

What this evaluates

Evaluates the capability of 2D and 3D convolutional neural networks to segment brain tumor sub-regions (enhancing tumor, whole tumor, tumor core) from multi-modal volumetric MRI scans. It specifically probes the effectiveness of ImageNet pretraining and architectural extensions on segmentation accuracy and robustness across benchmark and private clinical data.

Datasets

• BraTS 2020 — total ?; splits: train (-1), val (-1), test (-1)
• Syrian-Lebanese Hospital Clinical Dataset — total 5; splits: test (5)

Metrics

• Dice score (primary) — range: [0, 1]
  • 2 * |prediction ∩ ground_truth| / (|prediction| + |ground_truth|). Computed per voxel for each tumor sub-region (ET, WT, TC) and averaged across 5 runs.
• Hausdorff distance — range: other
  • Maximum distance between the boundaries of the prediction and ground truth masks. Reported in millimeters.

Input / output format

Input: Multi-modal volumetric MRI scans (T1c, T2, FLAIR) preprocessed to isotropic 1mm resolution, skull-stripped, and rigidly co-registered. During training, data is fed as patches of size (24x128x128) voxels with modalities treated as RGB channels.

Output: Volumetric segmentation masks predicting three tumor sub-regions: Enhancing Tumor (ET), Whole Tumor (WT), and Tumor Core (TC).

Scoring recipe

def dice_score(pred, gt):
    intersection = np.sum(pred * gt)
    return (2.0 * intersection) / (np.sum(pred) + np.sum(gt))

def hausdorff_distance(pred, gt):
    from scipy.spatial.distance import directed_hausdorff
    pred_boundary = np.argwhere(pred == 1)
    gt_boundary = np.argwhere(gt == 1)
    hd1 = directed_hausdorff(pred_boundary, gt_boundary)[0]
    hd2 = directed_hausdorff(gt_boundary, pred_boundary)[0]
    return max(hd1, hd2)

Common pitfalls

• MRI contrasts and resolutions vary significantly across clinical patients, requiring careful reverse-engineering of the BraTS preprocessing pipeline (DICOM to NIFTI, skull stripping, co-registration, resampling).
• The private clinical dataset lacks the T1 modality used in BraTS, forcing the use of T1c for both T1 and T1c channels, which may affect performance.
• Outliers are frequent in clinical evaluation, suggesting that mean/median scores alone may mask robustness issues.

Evidence (verbatim from paper)

Table 1 compares mean/median dice scores and Hausdorff distances for all segmentation labels and methods taken over 5 runs for the validation data and a single test run of the best-performing method.

Citation

@misc{wacker2019transfer,
  title={Transfer Learning for Brain Tumor Segmentation},
  author={Wacker et al. (2019)},
  year={2019},
  note={arXiv:1912.12452}
}

• arXiv: 1912.12452