Histopathology Grading Eval

---
name: histopathology-grading-eval
description: Evaluates deep learning models on binary classification of histopathology tissue tiles as malignant or benign. It probes the capacity of multi-stream architectures to capture diverse morphological and textural features for medical image grading. Use when the user wants to benchmark on CAMELYON16, Invasive Ductal Carcinoma (IDC), or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1806.08463
  bibtex_key: bidart2018triresnet
  confidence: high
---

# histopathology-grading-eval

> TriResNet: A Deep Triple-stream Residual Network for Histopathology Grading — Bidart et al. (2018) (arXiv:1806.08463, 2018)

## What this evaluates

Evaluates deep learning models on binary classification of histopathology tissue tiles as malignant or benign. It probes the capacity of multi-stream architectures to capture diverse morphological and textural features for medical image grading.

## Datasets

- **CAMELYON16** — total ?; splits: train (-1), val (-1), test (-1)
- **Invasive Ductal Carcinoma (IDC)** — total 198738; splits: train (-1), val (-1), test (-1)

## Metrics

- `sensitivity` — range: [0, 1]
  - True positive rate: TP / (TP + FN).
- `specificity` — range: [0, 1]
  - True negative rate: TN / (TN + FP).
- `accuracy` **(primary)** — range: [0, 1]
  - Overall correctness: (TP + TN) / (TP + TN + FP + FN).

## Input / output format

**Input**: RGB tissue image tiles extracted from whole slide images at 40x magnification (224x224 pixels for CAMELYON16, resized to 197x197 pixels for IDC).

**Output**: Binary class prediction: 'malignant' or 'benign'.

## Scoring recipe

```python
def compute_metrics(y_true, y_pred):
    tp = sum((t == 1 and p == 1) for t, p in zip(y_true, y_pred))
    tn = sum((t == 0 and p == 0) for t, p in zip(y_true, y_pred))
    fp = sum((t == 0 and p == 1) for t, p in zip(y_true, y_pred))
    fn = sum((t == 1 and p == 0) for t, p in zip(y_true, y_pred))
    sensitivity = tp / (tp + fn)
    specificity = tn / (tn + fp)
    accuracy = (tp + tn) / (tp + tn + fp + fn)
    return sensitivity, specificity, accuracy
```

## Common pitfalls

- Training data is artificially balanced by oversampling malignant tiles, which does not reflect the natural prevalence of cancer in histopathology slides.
- Evaluation is performed at the tile level rather than aggregating predictions to the whole-slide level, which may overestimate clinical diagnostic performance.
- No color normalization is applied, making results sensitive to staining variations across different scanners or institutions.

## Evidence (verbatim from paper)

> For both the CAMELYON16 and IDC datasets, we evaluated the performance of each tested network on their ability to grade tissue image tiles as either malignant or benign. For each network and dataset we evaluated the following three performance metrics on the test set: 1. $sensitivity\=TP/(TP+FN)$ 2. $specificity\=TN/(TN+FP)$ 3. $accuracy\=(TP+TN)/(TP+TN+FP+FN)$

## Citation

```bibtex
@misc{bidart2018triresnet,
  title={TriResNet: A Deep Triple-stream Residual Network for Histopathology Grading},
  author={Bidart et al. (2018)},
  year={2018},
  note={arXiv:1806.08463}
}
```

- arXiv: 1806.08463