Model Card for POCUS Aorta Segmentation using finetuned YOLOv8

Model Details

Model Description

This model is designed for Point of Care Ultrasound (POCUS) aorta segmentation, leveraging the finetuned YOLOv8 architecture. It aims to facilitate accurate and efficient aortic ultrasound image segmentation to aid medical professionals in diagnosing aortic conditions.

Developed by: [Sumit Pandey, Erik B. Dam and Kuan Fu Chen]
Funded by [optional]: [University of Copenhagen and Chang Gung Memorial Hospital Taiwan]
Shared by [optional]: [Sumit Pandey]
Model type: Convolutional Neural Network (CNN) for Object Detection and Segmentation
Language(s) (NLP): Not applicable
License: Apache-2.0
Finetuned from model [optional]: YOLOv8

Model Sources [optional]

Paper [optional]: [https://www.researchsquare.com/article/rs-4497019/v1]
Demo [optional]: [https://huggingface.co/spaces/sumit-ai-ml/Aorta-segmentation]

Uses

Direct Use

This model can be directly used to segment aortic structures in ultrasound images. It is intended for use by healthcare professionals and researchers in the field of medical imaging. However it will require further investigation.

Downstream Use [optional]

The model can be fine-tuned for other segmentation tasks in medical imaging or integrated into larger diagnostic systems.

Out-of-Scope Use

This model is not intended for use in non-medical imaging contexts or for segmentation tasks unrelated to aortic structures. It should not be used as a sole diagnostic tool without professional medical interpretation.

Bias, Risks, and Limitations

This model is trained on a specific dataset of aortic ultrasound images and may not generalize well to images from different sources or with different characteristics. The model might exhibit bias based on the demographic or technical attributes of the training data.

Recommendations

Users should be aware of potential biases and validate the model on their own data before deploying it in clinical settings. Regular updates and retraining with diverse datasets are recommended to maintain model performance and reduce bias.

How to Get Started with the Model

Use the following code to get started with the model:

# Example code to load and use the model
from ultralytics import YOLO

# Load the model
model = YOLO('path_to_your_finetuned_model.pt')

# Perform segmentation
results = model('path_to_ultrasound_image.jpg')

# Visualize results
results.show()

Training Details

Training Data

The model is trained on a dataset of annotated aortic ultrasound images. The dataset includes diverse cases to ensure the robustness of the model.

Training Procedure

The model was finetuned from YOLOv8 using the following procedure:

Preprocessing [optional]

Images were resized to a standard input size, and data augmentation techniques such as rotation, scaling, and flipping were applied to enhance model generalization.

Training Hyperparameters

Training regime: FP16 mixed precision
Batch size: 16
Epochs: around 100
Learning rate: 0.001

Speeds, Sizes, Times [optional]

Training duration: Approximately 2 hours
Model size: 6.42MB
Inference time per image: 0.05 seconds

Evaluation

Testing Data, Factors & Metrics

Testing Data

The testing dataset consists of a separate set of annotated aortic ultrasound images that were not seen by the model during training.

Factors

Evaluation was performed across different subpopulations, including variations in patient age, gender, and ultrasound device settings.

Metrics

Mean Intersection over Union (mIoU): 0.85
Dice Coefficient: 0.88
Precision: 0.87
Recall: 0.86

Results

The model demonstrated high accuracy and robustness across the test set, with consistent performance in various subgroups.

Summary

The finetuned YOLOv8 model for POCUS aorta segmentation achieved high precision and recall, making it suitable for clinical applications in aortic ultrasound imaging.

Model Examination [optional]

Interpretability techniques such as Grad-CAM were used to validate the model's focus on relevant aortic structures during segmentation.

Model Architecture and Objective

The model uses the YOLOv8 architecture, optimized for object detection and segmentation tasks in medical imaging.

Compute Infrastructure

Hardware

Training was performed on NVIDIA A100 GPUs with 40GB VRAM.

Software

Framework: PyTorch
Operating System: Ubuntu 20.04

Citation [optional]

BibTeX:

@misc{POCUS Aorta Segmentation,
  author = {Sumit Pandey, Erik B. Dam and Kuan Fu Chen, University of Copenhagen and Chang Gung Memorial Hospital Taiwan},
  title = {POCUS Aorta Segmentation using finetuned YOLOv8},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/sumit-ai-ml/POCUS_Aorta_segmentation}
}

APA:

Sumit Pandey, University of Copenhagen (2024). POCUS Aorta Segmentation using finetuned YOLOv8. Hugging Face. https://huggingface.co/sumit-ai-ml/POCUS_Aorta_segmentation

More Information [optional]

For more information, visit the model repository