---
dataset_info:
  features:
  - name: image_before
    dtype: image
  - name: image_after
    dtype: image
  - name: text_prompt
    dtype: string
  - name: label
    dtype: int32
  - name: change_description
    dtype: string
  - name: method
    dtype: string
  - name: pair_id
    dtype: string
  splits:
  - name: train
    num_bytes: 365953410
    num_examples: 630
  - name: validation
    num_bytes: 78418588
    num_examples: 135
  - name: test
    num_bytes: 78418588
    num_examples: 135
  download_size: 522486493
  dataset_size: 522790586
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-nc-sa-4.0
task_categories:
- image-classification
- image-text-to-text
- visual-question-answering
language:
- en
tags:
- medical
- dermatology
- melanoma
- skin-cancer
- temporal-change-detection
- synthetic-data
- medical-imaging
- multimodal
- medgemma
pretty_name: 'DermaCheck: Temporal Dermatoscopic Image Pairs'
size_categories:
- 1K<n<10K
---

# DermaCheck Temporal Pairs Dataset

## Dataset Description

Synthetic temporal image pairs for training MedGemma to detect changes in dermatoscopic images over time.

**Created for:** MedGemma Impact Challenge 2026 - Novel Task Prize (temporal change detection)

## Dataset Statistics

- **Total pairs:** 900
- **Train:** 630 pairs (70.0%)
- **Validation:** 135 pairs (15.0%)
- **Test:** 135 pairs (15.0%)

## Generation Methods

1. **Controlled Augmentation** (~50%): Original HAM10000 images augmented to simulate temporal evolution
- Size increase: 10-30%
- Border irregularity: Elastic transforms
- Color variation: HSV adjustments
- Based on clinical research: melanoma evolution patterns

2. **Natural Pairing** (~50%): Similar lesions from HAM10000 matched as temporal proxies
- Feature-based similarity matching
- Cosine similarity range: [0.6, 0.85]
- Same diagnosis category (melanoma focus)

3. **Unchanged Pairs** (~15%): Negative examples for balanced training
- Same image duplicated as before/after
- Label: 0 (no change)

## Dataset Structure

Each example contains:
- `image_before`: PIL Image (before timepoint)
- `image_after`: PIL Image (after timepoint)
- `text_prompt`: ABCDE-focused change detection question
- `label`: 0 (no change) or 1 (change detected)
- `change_description`: Educational explanation of changes
- `method`: Generation method (controlled_augmentation, natural_pairing, unchanged)
- `pair_id`: Unique identifier

## Source Data

Based on HAM10000 dataset:
- Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset. *Sci Data* 5, 180161 (2018).
- License: CC BY-NC-SA 4.0 (Non-commercial use)
- URL: https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000

## Intended Use

- Fine-tuning MedGemma for temporal change detection
- Educational tool development
- Research purposes only (not for clinical diagnosis)

## Citation

If you use this dataset, please cite both the original HAM10000 dataset and this derived work.

## License

CC BY-NC-SA 4.0 (Non-commercial use, per HAM10000 license)