Medical Seg Eval

medical-seg-eval

Learning with Limited Annotations: A Survey on Deep Semi-Supervised Learning for Medical Image Segmentation — Jiao et al. (2022) (arXiv:2207.14191, 2022)

What this evaluates

Evaluates deep learning models for medical image segmentation by quantifying region overlap and boundary alignment between predicted masks and expert-annotated ground truth.

Datasets

• LA (Left Atrium) — total 100; splits: train (80), test (20)
• Pancreas CT — total 82; splits: train (62), test (20)
• BraTS 2019 — total 335; splits: train (250), val (25), test (60)
• ACDC — total 100; splits: train (70), val (10), test (20)

Metrics

• DSC (primary) — range: [0, 1]
  • Dice Similarity Coefficient measures region overlap: DSC = 2|G ∩ S| / (|G| + |S|), where G is ground truth and S is segmentation result.
• IoU — range: [0, 1]
  • Intersection over Union (Jaccard) measures overlap: IoU = |G ∩ S| / |G ∪ S|.
• HD — range: other
  • Hausdorff Distance measures maximum boundary error: HD(∂G, ∂S) = max(max_x∈∂G min_y∈∂S ||x-y||_2, max_x∈∂S min_y∈∂G ||x-y||_2).
• 95HD — range: other
  • 95th percentile of the Hausdorff distances between boundary points, used to eliminate outlier influence.

Input / output format

Input: 3D or 2D medical imaging volumes (MRI or CT) with isotropic or fixed resolution, often intensity-normalized or clipped to specific Hounsfield Unit ranges.

Output: Binary or multi-class segmentation masks/labels corresponding to the input volume.

Scoring recipe

def compute_dsc(G, S):
    intersection = np.sum(G & S)
    return 2.0 * intersection / (np.sum(G) + np.sum(S))

def compute_iou(G, S):
    intersection = np.sum(G & S)
    union = np.sum(G | S)
    return intersection / union if union > 0 else 0.0

def compute_hd(G, S):
    dists = cdist(G, S)
    return max(np.max(np.min(dists, axis=1)), np.max(np.min(dists, axis=0)))

Common pitfalls

• DSC and IoU measure region overlap but ignore boundary errors, which can be critical for small lesions or thin structures.
• HD is highly sensitive to outliers; 95HD is preferred to mitigate the influence of small subsets of outlier boundary points.
• Semi-supervised splits vary across papers (e.g., 10% vs 20% labeled), making direct comparison difficult without standardized task settings.

Evidence (verbatim from paper)

For medical image segmentation tasks, Dice Similarity Coefficient (DSC) is a widely used metric to measure the region overlap ratio of the ground truth G and segmentation result S. Another similar metric IoU (or Jaccard) is used as an alternative for the evaluation. These two metrics are defined as follows: DSC = 2|G ∩ S|/(|G| + |S|), IoU = |G ∩ S|/|G ∪ S|.

Citation

@misc{jiao2022ssl4mis,
  title={Learning with Limited Annotations: A Survey on Deep Semi-Supervised Learning for Medical Image Segmentation},
  author={Jiao et al. (2022)},
  year={2022},
  note={arXiv:2207.14191}
}

• arXiv: 2207.14191