Mitotic Figure Detection Eval

---
name: mitotic-figure-detection-eval
description: Evaluates the ability of object detection models to accurately localize and classify mitotic figures in histopathology images across varying stain domains and scales. It probes model robustness to domain shift and architectural trade-offs between anchor-based and anchor-free detection paradigms. Use when the user wants to benchmark on MIDOG 2025, or asks about evaluating this task. Reports F1 Score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2509.02957
  bibtex_key: [REDACTED]
  confidence: high
---

# mitotic-figure-detection-eval

> Ensemble YOLO Framework for Multi-Domain Mitotic Figure Detection in Histopathology Images — Kelam et al. (2025) (arXiv:2509.02957, 2025)

## What this evaluates

Evaluates the ability of object detection models to accurately localize and classify mitotic figures in histopathology images across varying stain domains and scales. It probes model robustness to domain shift and architectural trade-offs between anchor-based and anchor-free detection paradigms.

## Datasets

- **MIDOG 2025** — total ?; splits: internal_validation (-1), test (-1)

## Metrics

- `F1 Score` **(primary)** — range: percent
  - Harmonic mean of precision and recall: F1 = 2 * (Precision * Recall) / (Precision + Recall). Computed per image or dataset level based on true positive, false positive, and false negative bounding box matches.
- `Precision` — range: percent
  - Ratio of correctly detected mitotic figures to all predicted detections: Precision = TP / (TP + FP).
- `Recall` — range: percent
  - Ratio of correctly detected mitotic figures to all ground-truth instances: Recall = TP / (TP + FN).

## Input / output format

**Input**: Histopathology tissue images (patches or whole-slide regions) containing mitotic figures.

**Output**: Bounding box coordinates and class labels for each detected mitotic figure per image.

## Scoring recipe

```python
def compute_metrics(predictions, gold_boxes, iou_threshold=0.5):
    tp, fp, fn = 0, 0, 0
    matched_gold = set()
    for pred in predictions:
        best_iou, best_idx = 0, -1
        for i, gold in enumerate(gold_boxes):
            if i not in matched_gold:
                iou = calculate_iou(pred.box, gold.box)
                if iou > best_iou:
                    best_iou, best_idx = iou, i
        if best_iou >= iou_threshold:
            tp += 1
            matched_gold.add(best_idx)
        else:
            fp += 1
    fn = len(gold_boxes) - len(matched_gold)
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    f1 = 2 * (precision * recall) / (precision + recall) if (precision + recall) > 0 else 0.0
    return precision, recall, f1
```

## Common pitfalls

- Domain shift from varying histological stain colors across labs can cause anchor-based models to miss atypical mitotic figures, artificially lowering recall if not augmented.
- Small or overlapping mitotic figures are frequently under-detected; evaluation must account for size variation and use appropriate IoU thresholds to avoid penalizing slightly misaligned boxes.

## Evidence (verbatim from paper)

> On the preliminary MIDOG 2025 leaderboard, our ensembled method was ranked 5th with an F1 score of 0.7923, precision of 0.7357, and recall of 0.8583 (Table [1]). These results are consistent with internal validation, underscoring that ensembling balances YOLOv5’s precision bias with YOLOv8’s recall-oriented design.

## Citation

```bibtex
@misc{kelam2025ensemble,
  title={Ensemble YOLO Framework for Multi-Domain Mitotic Figure Detection in Histopathology Images},
  author={Kelam et al. (2025)},
  year={2025},
  note={arXiv:2509.02957}
}
```

- arXiv: 2509.02957