Retuve YOLO Segmentation AI Plugin

For more information on Retuve, see https://github.com/radoss-org/retuve

This codebase has the AI Plugin for Retuve, which uses Radiopedia data from The Open Hip Dataset to train.

The model weights are strictly under the combined terms of the CC BY-NC-SA 3.0 license, and the AGPL Licence. This is because the model is trained on Radiopedia Data, which is under the CC BY-NC-SA 3.0 license, and the YOLO ultralytics codebase is under the AGPL Licence.

This means that you cannot use this codebase for any commercial purposes, you must attribute Radiopedia for the data used to train the model, and you must obide by the terms of the AGPL Licence.

The codes dual licences are in the LICENSE file and the LICENSE2 file.

Initial Results

For a more detailed look at the validation results, see:

• Ultrasound Validation Results
• X-ray Validation Results

Detailed Performance Metrics

Ultrasound Validation Results (Limited Cases)

Metric	Results
Alpha Angle ICC	0.86 (95% CI -0.07-0.81)
Coverage ICC	0.92 (95% CI 0.84-0.96)

Ultrasound validation results for alpha angle and coverage measurements

X-ray Validation Results

Metric	Left Side	Right Side
ICC Acetabular Index	0.860 (95% CI 0.830-0.880)	0.845 (95% CI 0.810-0.870)
ICC Wilberg Index	0.891 (95% CI 0.860-0.910)	0.902 (95% CI 0.860-0.930)

X-ray validation results for acetabular index and Wilberg angle measurements

Classification Performance

For X-ray DDH classification, Retuve demonstrated strong performance in distinguishing Grade 1 IHDI from Grades 2, 3, and 4:

Classification Task	F1 Score	Recall	Precision
Grade 1 vs. Grades 2-4 IHDI	0.940	0.914	0.967
Per-Class (All Grades)	0.593	0.570	0.637

Classification performance for DDH grading on X-ray images

Note: For the Grade 1 vs. Grades 2-4 classification analysis, cases where Retuve returned a result of "0" were logically classified as IHDI Grade 2 or higher, as a "0" result represents a Retuve processing error and indicates the system's inability to confidently classify the case as normal (Grade 1).

F1 vs Confidence Score Analysis

The F1 vs confidence score plots shown in the detailed validation results help users select optimal confidence thresholds for their specific use case. The F1 score balances precision (accuracy of positive predictions) and recall (ability to find all positive cases), providing a single metric to evaluate model performance across different confidence levels.

Why this matters for inference:

• Screening applications: Lower thresholds (0.2-0.4) maximize recall to avoid missing cases
• Diagnostic applications: Higher thresholds (0.5-0.7) maximize precision for confident diagnoses
• Research applications: Mid-range thresholds (0.3-0.5) provide balanced performance

Users should select confidence thresholds based on their clinical priorities: whether it's more important to catch all potential cases (high recall) or to minimize false positives (high precision).

Training Parameters Summary

Parameter	Ultrasound	X-ray Segmentation	X-ray Pose
Model Architecture	YOLOv11n-seg	YOLOv11n-seg	YOLOv11n-pose
Task	Segmentation	Segmentation	Pose Estimation
Epochs	100	200	100
Batch Size	4	16	16
Image Size	500px	800px	800px
Learning Rate	0.01	0.01	0.01
Data Augmentation	Minimal	Extensive	Moderate
Key Augmentations	HSV (0.1), Rotation (5°)	HSV (0.7), Mosaic, AutoAug	HSV (0.7), AutoAug, Flip
Device	cuda:4	cuda:1	Auto

Training configurations optimized for each modality's specific requirements and dataset characteristics

Installation

From PyPI

To install the plugin from PyPI:

uv pip install retuve-yolo-plugin

From Source

To install the plugin from source:

pip install git+https://github.com/radoss-org/retuve-yolo-plugin.git

Development Installation

For development, clone the repository and install with uv:

git clone https://github.com/radoss-org/retuve-yolo-plugin.git
cd retuve-yolo-plugin
uv sync --dev

Running Tests and Formatting

# Run tests
uv run pytest -vv -n 4 ./tests

# Format code
uv run black .

# Check formatting
uv run black --check .

Example Usage

Please see https://github.com/radoss-org/retuve/tree/main/examples for more examples. This is purely meant to illustrate how to use the plugin.

import pydicom
from retuve.defaults.hip_configs import default_US
from retuve.funcs import analyse_hip_3DUS
from retuve.testdata import Cases, download_case

from retuve_yolo_plugin.ultrasound import yolo_predict_dcm_us

# Get an example case
dcm_file = download_case(Cases.ULTRASOUND_DICOM)[0]

default_US.device = "cpu"

dcm = pydicom.dcmread(dcm_file)

hip_datas, *_ = analyse_hip_3DUS(
    dcm,
    keyphrase=default_US,
    modes_func=yolo_predict_dcm_us,
    modes_func_kwargs_dict={},
)

print(hip_datas)

Attribution

We give full attribution to the authors that made this effort possible on Radiopedia. The list of these authors can be found here.

License

The codes dual licences are in the LICENSE file and the LICENSE2 file.

If you are interested in a less-restritive licence, the first step is to contact Radiopedia for a special licence to use all the data this model is trained on. That list can be found here.

The 2nd step is to contact YOLO Ultralytics for a commercial licence for their codebase. That process is described here.

RadOSS will then consider providing you a commercial licence for this plugin at no charge. Please contact us at [email protected] when you have obtained the licence from Radiopedia and YOLO Ultralytics.