This repository contains the code and documentation for the master thesis project, which focuses on developing a deep learning model to correct attenuation and scatter corrections without anatomical inforamtion from CT, also detect and correct artefacts in 68Ga-PET imaging. In this thesis, we have demonstrated the efficacy of an Integrated multi-center Dynamic Unet DL framework for artifact detection and correction in PET imaging of 68Ga-labelled compounds. The approach leverages large datasets from multiple centers, to achive a universal model for same radiotracer but different centers. Through the incorporation of transfer learning concepts, we have developed radiotracer-specific models that significantly outperform those based on specific radiotracers, thereby addressing a major limitation in the field of medical imaging. Our model effectively detected and corrected artifacts. This enhancement is vital for making therapeutic decisions in the field of oncology, where PET imaging plays a central role in diagnosing, planning treatments, and evaluating responses. By using Dyn-Unet architecture and other advanced DL techniques, our method has not only improved image quality but also greatly decreased the appearance of common artifacts like halo and mismatch artifacts, especially in 68Ga-PET imaging. The effective implementation of our models in different centers highlights their resilience and flexibility, which are essential for general acceptance in clinical settings.
In this study we tried to take a step into the problematic field of correction in PET imaging artefacts, with especially high-prevalence ones: mismatch and halo artefacts in 68Ga PET imaging. The aim of this paper is to look at several deep learning models and methodologies to design a multi-centre model that allows un-direct data sharing at each centre due to some demerits of conventional deep learning techniques. We will use our approach to estimate and compare the performance of models under both strategies within different radio tracer distributions and multi-centre data environments. This research will aim to be a demonstration of the possibility, as well as the superior performance, of the deep learning models for real clinical settings, which will potentially set a new standard of CT-free PET imaging that enhances diagnostic accuracy while minimising radiation exposure and procedural complexity.
- Integrated Multi-Center Model (IMCM): Combines data from multiple centers to enhance model generalization.
- Anatomy-Dependent Correction Model (ADCM): Focuses on correcting attenuation and scatter based on anatomical features.
- Tuned Transfer Learning Model (TL-MC): Addresses the variability across different radiotracers.
src/: Source code for the deep learning models and utility functions.data/: Sample data files (note: due to privacy, actual patient data are not included).Archive/: Archived notbooks nessasary for further investigation.Analysis/: Archived notbooks about data pre processing, statistical testing and so on.Other_Models/: Archived models notbooks we could'nt get good results or needing for further progress.info/: Documentation and additional notes on the project.Results/: Analysis output, plots, model structure and images used in the thesis for demonstration, such as artefact corrections.
Ensure you have Python 3.8+ installed, then run:
pip install -r requirements.txtUsage To train the model:
python src/main.pyTo evaluate the model:
python Inference.ipynbTo use the application:
streamlit run dashboard.pyThis project achieved notable improvements in artefact detection and attenuation/scatter correction. Refer to the results/ directory for detailed performance metrics and visual results.
This section showcases the results of the artifact correction applied using the IMCM model:
This project is licensed under the MIT License - see the LICENSE.md file for details.
Hanze University of applied science, Groningen, The Netherlands
Author: Sama Shahpouri
Email: [[email protected]]