Mpunet Tumor Seg Eval

---
name: mpunet-tumor-seg-eval
description: Evaluates a modified 2D U-Net's ability to segment pediatric and adult brain tumors in MRI scans by leveraging multi-planar data augmentation to learn 3D volumetric representations. It probes the model's generalization across diverse tumor types, anatomical variations, and imaging scenarios. Use when the user wants to benchmark on Pediatrics Tumor Challenge (PED), Brain Metastasis Challenge (MET), Sub-Sahara-Africa Adult Glioma Challenge (SSA), or asks about evaluating this task. Reports Dice Score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2401.06499
  bibtex_key: pandey2024mpunet
  confidence: high
---

# mpunet-tumor-seg-eval

> Fully Automated Tumor Segmentation for Brain MRI data using Multiplanner UNet — Pandey et al. (2024) (arXiv:2401.06499, 2024)

## What this evaluates

Evaluates a modified 2D U-Net's ability to segment pediatric and adult brain tumors in MRI scans by leveraging multi-planar data augmentation to learn 3D volumetric representations. It probes the model's generalization across diverse tumor types, anatomical variations, and imaging scenarios.

## Datasets

- **Pediatrics Tumor Challenge (PED)** — total ?; splits: train (-1), val (-1), test (-1)
- **Brain Metastasis Challenge (MET)** — total ?; splits: train (-1), val (-1), test (-1)
- **Sub-Sahara-Africa Adult Glioma Challenge (SSA)** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `Dice Score` **(primary)** — range: [0, 1]
  - Quantifies overlap between predicted segmentation mask and ground truth mask. Calculated as 2 * |A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth masks respectively.

## Input / output format

**Input**: 3D brain MRI volumes (using first 3 of 4 available modalities), processed with multi-planar data augmentation (rotations along various axes to generate multiple 2D perspectives).

**Output**: Binary segmentation masks for three tumor subregions: enhancing tumor (ET), tumor core (TC), and whole tumor (WT).

## Scoring recipe

```python
def dice_score(pred, gt):
    intersection = np.sum(pred * gt)
    union = np.sum(pred) + np.sum(gt)
    return (2.0 * intersection) / (union + 1e-6)
```

## Common pitfalls

- The paper reports Dice scores separately for ET, TC, and WT but does not specify a single aggregate metric for overall model comparison.
- The cross-validation description ambiguously combines test and validation roles, making the exact train/val/test split ratio unclear.
- Only 3 out of 4 available MRI modalities are used to reduce GPU memory, potentially underreporting maximum achievable performance.

## Evidence (verbatim from paper)

> To evaluate the precision of our generated masks, we utilized the Dice Score matrix as the primary evaluation metric across all three distinct classes: ET (enhancing tumor), TC (tumor core), and WT (whole tumor). ... The evaluation was conducted using three distinct challenges: the Pediatrics Tumor Challenge (PED), the Brain Metastasis Challenge (MET), and the Sub-Sahara-Africa Adult Glioma Challenge (SSA). ... To ensure reliable evaluation, a rigorous 3-fold cross-validation technique was employed.

## Citation

```bibtex
@misc{pandey2024mpunet,
  title={Fully Automated Tumor Segmentation for Brain MRI data using Multiplanner UNet},
  author={Pandey et al. (2024)},
  year={2024},
  note={arXiv:2401.06499}
}
```

- arXiv: 2401.06499