Oesophageal Adenocarcinomas Eval

---
name: oesophageal-adenocarcinomas-eval
description: Evaluates the generalization capability of deep learning architectures on histopathology image classification, specifically probing how model capacity affects overfitting on sparse, high-resolution medical imaging data. Use when the user wants to benchmark on Oesophageal Adenocarcinomas Dataset, or asks about evaluating this task. Reports validation F1 score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2506.16631
  bibtex_key: [REDACTED]
  confidence: high
---

# oesophageal-adenocarcinomas-eval

> Overfitting in Histopathology Model Training: The Need for Customized Architectures — Alfasly et al. (2025) (arXiv:2506.16631, 2025)

## What this evaluates

Evaluates the generalization capability of deep learning architectures on histopathology image classification, specifically probing how model capacity affects overfitting on sparse, high-resolution medical imaging data.

## Datasets

- **Oesophageal Adenocarcinomas Dataset** — total ?; splits: train (-1), val (-1)

## Metrics

- `validation F1 score` **(primary)** — range: [0, 1]
  - The harmonic mean of precision and recall computed on the validation set. The paper reports this as the headline metric for model generalization, with higher values indicating better performance.

## Input / output format

**Input**: High-resolution histopathology images of oesophageal tissue.

**Output**: Predicted class labels for each image instance.

## Scoring recipe

```python
def compute_f1(preds, gold):
    tp = sum(p == g == 1 for p, g in zip(preds, gold))
    fp = sum(p == 1 and g == 0 for p, g in zip(preds, gold))
    fn = sum(p == 0 and g == 1 for p, g in zip(preds, gold))
    prec = tp / (tp + fp) if (tp + fp) > 0 else 0
    rec = tp / (tp + fn) if (tp + fn) > 0 else 0
    return 2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0
```

## Common pitfalls

- High training F1 (0.96-0.97) paired with significantly lower validation F1 (0.56-0.77) indicates severe overfitting, which is the primary phenomenon being measured rather than absolute performance.
- Model capacity does not linearly correlate with better validation performance; intermediate architectures (e.g., ResNet-50, ViT depth 16) often outperform larger variants on this dataset.
- F1 score is reported without specifying macro vs. micro averaging, which can lead to ambiguity in multi-class settings.

## Evidence (verbatim from paper)

> Training F1 scores remained consistently high (0.96-0.97) in all ResNet variants, while validation F1 scores showed minimal fluctuation (0.73-0.77), indicating significant overfitting. Even SimpleCNN with only six layers achieved comparable validation performance (F1 score 0.75) to more complex architectures, suggesting that additional model capacity does not translate to better generalization.

## Citation

```bibtex
@misc{alfasly2025overfitting,
  title={Overfitting in Histopathology Model Training: The Need for Customized Architectures},
  author={Alfasly et al. (2025)},
  year={2025},
  note={arXiv:2506.16631}
}
```

- arXiv: 2506.16631