README.md

---
license: agpl-3.0
library_name: ultralytics
tags:
  - object-detection
  - medical-imaging
  - colonoscopy
  - polyp-detection
  - endoscopy
pipeline_tag: object-detection
---

# TRUE-Colon — RT-DETR (REAL-Colon, seed 42)

Official weights for **TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time
Polyp Detection** (MICCAI 2026 EndoLINA Workshop). This repo holds the **`rtdetr`**
detector at **seed 42**, trained on REAL-Colon: 60 complete, unedited colonoscopy
procedures rather than curated lesion-centric clips.

TRUE-Colon is an evaluation protocol first. Its finding is that detectors trained on
curated benchmarks collapse on full procedures, while procedure-trained models like this
one suppress non-polyp content and still transfer back to curated data. The detector itself
is an off-the-shelf RT-DETR: what is released here is a baseline trained under the
paper's protocol, not a new architecture.

This is one of three seeds. Reporting a single seed of this model in isolation would
overstate its precision; the paper aggregates all three.

## Model Details

- **Developed by:** Sebastian Doerrich, Andreas Franz Schwab (xAILab, University of Bamberg)
- **Funded by:** Hightech Agenda Bayern (HTA) of the Free State of Bavaria, Germany
- **Model type:** Single-class object detector (`lesion`)
- **License:** AGPL-3.0, inherited from Ultralytics. This is **not** the MIT license of the
  `true_colon` evaluation package (© Sebastian Doerrich, Andreas Franz Schwab). Using these
  weights in a product carries the AGPL's source-availability obligations.
- **Framework:** Ultralytics

### Model Sources

- **Paper:** https://arxiv.org/abs/2608.13711
- **Repository:** https://github.com/sdoerrich97/true-colon
- **Package:** https://pypi.org/project/true-colon/

### Architecture and training

| | |
|---|---|
| Backbone | RT-DETR |
| Input | 640 x 640 |
| Classes | 1 (`lesion`) |
| Training data | REAL-Colon, patient-level split, 86.47% lesion-free training frames |
| Seed | 42 |
| Batch size | 66 |
| Optimizer | SGD, lr0 0.01, momentum 0.9, weight decay 5e-4 (resolved from `optimizer=auto`) |
| Schedule | 100 epochs, early stopping patience 10, 3-epoch warm-up |
| Initialization | COCO-pretrained, as shipped by Ultralytics |
| Ultralytics version | 8.3.232 |

This is the configuration this checkpoint was trained with. The paper states the recipe in
one sentence and otherwise defers to the frameworks' defaults, so the values it does not
state — the optimizer, learning rate, momentum, weight decay and warm-up — come from the
training run itself.

> **On the optimizer.** The run recorded `optimizer: auto`, which is the framework
> default the paper keeps. Ultralytics resolves that at runtime: SGD when the iteration
> budget exceeds 10,000, AdamW otherwise. This run reaches 2,585,900 iterations, so
> it trained with **SGD**, at `lr0=0.01` and `momentum=0.9` — `auto` sets both, so the
> run's own `momentum` entry does not apply.

## Uses

### Direct use

Research on polyp detection and on evaluation methodology for video-based computer-aided
detection. The intended use is reproducing and extending the TRUE-Colon benchmark.

### Out-of-scope use

**This is not a medical device and must not be used for clinical decision-making.** The
paper's own conclusion is that performance under realistic full-procedure conditions
remains insufficient for reliable clinical deployment. It has not been validated
prospectively, cleared by any regulator, or tested against the population, scope hardware
or bowel-preparation quality of any particular site.

## Bias, risks and limitations

- Trained on 60 procedures from 4 institutions; endoscope hardware, imaging mode and
  patient population all shift performance.
- Detection is dominated by medium and large lesions; average precision on small lesions
  is near zero.
- Two histological subtypes (SSL, TSA) appear only in the test split, and sessile serrated
  lesions have the highest miss rates.
- Frame-level, with no temporal model: each frame is scored independently.

## How to get started

```python
from ultralytics import YOLO  # or RTDETR for the RT-DETR variant
from huggingface_hub import hf_hub_download

weights = hf_hub_download("sdoerrich97/true_colon_rtdetr_realcolon_s42", "model.pt")
model = YOLO(weights)
results = model.predict("frame.jpg", conf=0.30)
```

Then evaluate under the protocol:

```python
import true_colon as tc

evaluator = tc.create_evaluator(dataset="realcolon", id_source="pred_string")
result = evaluator.evaluate_paths("gt.json", "predictions.json")
```

## Evaluation

Evaluate with the TRUE-Colon protocol rather than COCO mAP alone: the point of the
benchmark is that localization accuracy alone does not predict deployment behaviour. The
paper reports this architecture at a matched false-alert operating point of
**tau* = 0.30** on REAL-Colon, targeting a frame-level false-positive rate of 4-5%.

For the full result tables, see the paper. Per-seed numbers are not reproduced on this
card, because a single seed's values are not the quantity the paper reports.

## Citation

```bibtex
@article{doerrich2026truecolon,
  title={TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection},
  author={Sebastian Doerrich and Andreas Franz Schwab and Francesco {Di Salvo} and Shyam Nandan Rai and Hanh Huyen My Nguyen and Christian Ledig},
  year={2026},
  eprint={2608.13711},
  archivePrefix={arXiv},
  primaryClass={eess.IV},
  url={https://arxiv.org/abs/2608.13711},
}
```

## Model Card Contact

For questions or issues, please open an issue in the
[GitHub repository](https://github.com/sdoerrich97/true-colon), or contact
sebastian.doerrich@uni-bamberg.de.