PULASki Open-Source Biomedical Image Segmentation Tool - Efficiently Tackle Small Datasets and Class Imbalance

Pulaski ProbUNet3D Base VSeg

Developed by soumickmj

PULASki is a computationally efficient biomedical image segmentation tool that accurately captures variability in expert annotations, particularly suitable for small datasets and class imbalance issues.

Image Segmentation

Safetensors

Open Source License:Apache-2.0 #7T MRI vessel segmentation #Probabilistic annotation learning #Small dataset optimization

Downloads 14

Release Time : 9/3/2024

Model Overview

This model is a 3D medical image segmentation model based on conditional variational autoencoder architecture (Probabilistic UNet), specifically designed for vessel segmentation in 7T Time-of-Flight MR angiography volumetric data.

Model Features

Handling annotation variability

Accurately captures high variability in multi-expert annotations, improving clinical applicability of segmentation results.

Small dataset efficiency

Achieves good segmentation performance even with scarce annotated small datasets.

Class imbalance optimization

Uses improved loss function (Focal Tversky loss) to significantly enhance learning capability for imbalanced data.

Uncertainty quantification

Provides probabilistic segmentation results, avoiding misleading overconfident predictions.

Model Capabilities

3D medical image segmentation

Vessel structure recognition

Probabilistic prediction output

Processing 7T Time-of-Flight MR data

Use Cases

Medical image analysis

Cerebrovascular segmentation

Used for cerebrovascular structure segmentation in 7T Time-of-Flight MR angiography

Accurately identifies vascular structures and quantifies prediction uncertainty

Clinical decision support

Provides physicians with probabilistic segmentation results of vascular structures to assist diagnosis and treatment planning

Reduces bias from single annotations and improves clinical decision reliability

🚀 PULASki_ProbUNet3D_Base_VSeg

In the field of medical imaging, this project offers a computationally efficient generative tool for biomedical image segmentation, addressing challenges like annotation variability and data scarcity.

🚀 Quick Start

In the domain of medical imaging, many supervised learning based methods for segmentation face several challenges such as high variability in annotations from multiple experts, paucity of labelled data and class imbalanced datasets. These issues may result in segmentations that lack the requisite precision for clinical analysis and can be misleadingly overconfident without associated uncertainty quantification.

We proposed the PULASki as a computationally efficient generative tool for biomedical image segmentation that accurately captures variability in expert annotations, even in small datasets. Our approach makes use of an improved loss function based on statistical distances in a conditional variational autoencoder structure (Probabilistic UNet), which improves learning of the conditional decoder compared to the standard cross-entropy particularly in class imbalanced problems.

✨ Features

Accurate Variability Capture: Captures variability in expert annotations, even in small datasets.
Improved Loss Function: Utilizes an improved loss function based on statistical distances in a conditional variational autoencoder structure, enhancing learning in class - imbalanced problems.

📚 Documentation

Model Details

It was introduced in PULASki: Learning inter-rater variability using statistical distances to improve probabilistic segmentation by Soumick Chatterjee, Franziska Gaidzik, Alessandro Sciarra, Hendrik Mattern, Gábor Janiga, Oliver Speck, Andreas Nürnberger, Sahani Pathiraja.

Model Description

Property	Details
Developed by	Dr Soumick Chatterjee
Model Type	3D Probabilistic UNet (PULASki's baseline), trained (without a distribution distance loss) with Focal Tversky loss (FTL)
Task	Probabilistic vessel segmentation in 7T MRA - ToF volumes
Training Data	7T MRA - ToF volumes, details mentioned in Sec. 4.1 of https://arxiv.org/pdf/2312.15686

Model Sources

Repository: https://github.com/soumickmj/PULASki
Paper: https://arxiv.org/abs/2312.15686

Citation

If you use this approach in your research or use codes from this repository or these weights, please cite the following in your publications:

BibTeX:

@article{chatterjee2023pulaski,
  title={PULASki: Learning inter-rater variability using statistical distances to improve probabilistic segmentation},
  author={Chatterjee, Soumick and Gaidzik, Franziska and Sciarra, Alessandro and Mattern, Hendrik and Janiga, G{\'a}bor and Speck, Oliver and N{\"u}rnberger, Andreas and Pathiraja, Sahani},
  journal={arXiv preprint arXiv:2312.15686},
  year={2023}
}

APA: Chatterjee, S., Gaidzik, F., Sciarra, A., Mattern, H., Janiga, G., Speck, O., Nuernberger, A., & Pathiraja, S. (2023). PULASki: Learning inter-rater variability using statistical distances to improve probabilistic segmentation. arXiv preprint arXiv:2312.15686.

📄 License

This project is licensed under the Apache - 2.0 license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご