|
|
--- |
|
|
license: apache-2.0 |
|
|
pipeline_tag: image-segmentation |
|
|
--- |
|
|
|
|
|
# nnActive: Toothfairy2 Queries and Analysis Results |
|
|
|
|
|
This repository contains the queries and analysis results for the **Toothfairy2** dataset, as presented in the papers: |
|
|
|
|
|
* **Finally Outshining the Random Baseline: A Simple and Effective Solution for Active Learning in 3D Biomedical Imaging** (TMLR 2026). [[Paper]](https://huggingface.co/papers/2601.13677) |
|
|
* **nnActive: A Framework for Evaluation of Active Learning in 3D Biomedical Segmentation** (TMLR 2025). [[Paper]](https://arxiv.org/abs/2511.19183) |
|
|
|
|
|
Official Code: [GitHub - MIC-DKFZ/nnActive](https://github.com/MIC-DKFZ/nnActive/tree/nnActive_v2) |
|
|
|
|
|
## Summary |
|
|
Active learning (AL) aims to reduce annotation costs in 3D biomedical image segmentation by selecting the most informative samples. However, standard uncertainty-based AL methods often struggle to outperform random baselines due to class imbalance and query redundancy. |
|
|
|
|
|
This work introduces **ClaSP PE** (Class-stratified Scheduled Power Predictive Entropy), a simple and effective query strategy that addresses these limitations through: |
|
|
1. **Class-stratified querying**: Ensures coverage of underrepresented structures. |
|
|
2. **Scheduled Power Noising**: Enforces query diversity in early-stage AL and encourages exploitation later. |
|
|
|
|
|
The results in this repository were generated using the **nnActive** framework, an open-source evaluation framework for 3D medical AL built on top of nnU-Net. |
|
|
|
|
|
## Artifact Information |
|
|
The repository includes: |
|
|
- `loop_XXX.json`: Lists of specific 3D patches selected for annotation in each Active Learning cycle. |
|
|
- `dataset.json`: Metadata and label configuration for the Toothfairy2 experiment. |
|
|
- `auc.json` & `summary.json`: Detailed performance metrics and Dice score statistics. |
|
|
- `config.json`: Experimental parameters for reproducibility. |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use the **nnActive** framework or these results, please cite: |
|
|
|
|
|
```bibtex |
|
|
@article{luth2025nnactive, |
|
|
title={nnActive: A Framework for Evaluation of Active Learning in 3D Biomedical Segmentation}, |
|
|
author={Carsten T. L{\"u}th and Jeremias Traub and Kim-Celine Kahl and Till J. Bungert and Lukas Klein and Lars Kr{\"a}mer and Paul F Jaeger and Fabian Isensee and Klaus Maier-Hein}, |
|
|
journal={Transactions on Machine Learning Research}, |
|
|
issn={2835-8856}, |
|
|
year={2025}, |
|
|
url={https://openreview.net/forum?id=AJAnmRLJjJ}, |
|
|
} |
|
|
``` |
|
|
|
|
|
If you use **ClaSP PE**, please cite: |
|
|
|
|
|
```bibtex |
|
|
@article{luth2026finally, |
|
|
title={Finally Outshining the Random Baseline: A Simple and Effective Solution for Active Learning in 3D Biomedical Imaging}, |
|
|
author={Carsten T. L{\"u}th and Jeremias Traub and Kim-Celine Kahl and Till J. Bungert and Lukas Klein and Lars Kr{\"a}mer and Paul F Jaeger and Klaus Maier-Hein and Fabian Isensee}, |
|
|
journal={Transactions on Machine Learning Research}, |
|
|
issn={2835-8856}, |
|
|
year={2026}, |
|
|
url={https://openreview.net/forum?id=UamXueEaYW}, |
|
|
} |
|
|
``` |
