Explainseg Segmentation Eval

explainseg-segmentation-eval

No Masks Needed: Explainable AI for Deriving Segmentation from Classification — Ma et al. (2025) (arXiv:2508.04534, 2025)

What this evaluates

Evaluates the ability of an XAI-driven classification model to generate clinically meaningful segmentation masks without pixel-level annotations. It probes spatial coherence, boundary precision, and generalization across diverse medical imaging modalities (mammography, histopathology, endoscopy).

Datasets

• CBIS-DDSM — total ?; splits: train (1276), val (234)
• NuInsSeg — total 665; splits: train (600), val (65)
• Kvasir-SEG — total 1000; splits: train (1852), val (254)

Metrics

• mIoU — range: percent
  • Mean Intersection over Union across all classes: average of (intersection / union) per class.
• Dice (primary) — range: percent
  • Dice coefficient: 2 * |A ∩ B| / (|A| + |B|), measuring overlap between predicted and ground truth masks.

Input / output format

Input: Normalized medical images (mammography, histopathology tiles, or colonoscopy patches) passed through a fine-tuned Vision Transformer (ViT) classification backbone.

Output: Binary segmentation masks derived from Integrated Gradients relevance heatmaps, optionally fused with ViT features, and refined via Normalized Cut (NCut) or morphological operations.

Scoring recipe

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

def compute_miou(pred, gt, classes=2):
    ious = []
    for c in range(classes):
        p = (pred == c)
        g = (gt == c)
        inter = np.logical_and(p, g).sum()
        union = np.logical_or(p, g).sum()
        ious.append(inter / union if union > 0 else 0.0)
    return np.mean(ious) * 100

Common pitfalls

• Using ground truth segmentation masks during the classification network fine-tuning stage, which violates the paper's annotation-free premise.
• Directly comparing with fully supervised segmentation baselines without acknowledging the lack of pixel-level labels during training.
• Computing normalization statistics on the full dataset instead of solely the training subset, causing data leakage.

Evidence (verbatim from paper)

ExplainSeg (XAI & NCut) consistently achieves the best or near-best segmentation performance across all three datasets. It yields the highest mIoU (31.2%) and Dice score (43.7%) on CBIS-DDSM, the mammography dataset that poses substantial challenges due to low contrast and subtle boundaries.

Citation

@misc{ma2025nomasks,
  title={No Masks Needed: Explainable AI for Deriving Segmentation from Classification},
  author={Ma et al. (2025)},
  year={2025},
  note={arXiv:2508.04534}
}

• arXiv: 2508.04534