Stories by Ibrahim Abdelazim on Medium

When working with Siemens MRI scanners, you may encounter the term' Phoenix Protocol'. This protocol is embedded within DICOM files in what so called CSA Headers and contains a wealth of information about the acquisition setup, including parameters such as sequence file name, echo time, repetition time, slice information, and more.
Understanding how to extract and parse this protocol can be very useful for medical imaging research, sequence reproducibility, or even debugging data acquisition pipelines. In this article, I’ll explain what the Phoenix Protocol is, how it’s stored in Siemens DICOM files, and how you can read it with Python.
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*QpmycD_cw3WEikDhkAfJvw.jpeg"></figure><h3>What is the Phoenix Protocol?</h3>
The Phoenix Protocol is Siemens’ way of storing a full snapshot of the MRI acquisition parameters. Think of it as a human-readable script describing how the MRI scan was configured. It typically contains information like:
<ul>
<li>Sequence file name</li>
<li>TR, TE, and TI values</li>
<li>Slice orientation and positioning</li>
<li>Parallel imaging settings</li>
<li>RF pulse shapes and timing</li>
</ul>
This protocol ensures that the scan can be reconstructed or replicated later — a crucial part of quality assurance and reproducibility in MRI research.
<h3>What is a CSA Header/Tag?</h3>
Siemens embeds this private data inside DICOM files using CSA headers (Common Siemens Attributes). These aren’t part of the standard DICOM attributes but are stored in private tags. It is a binary-encoded key–value structure.
These headers are stored as two <a href="http://dicom.nema.org/medical/dicom/current/output/html/part05.html#sect_7.8">private data elements</a>:
<ul>
<li>(0029, 1010) — CSA Image Header Info</li>
<li>(0029, 1020) — CSA Series Header Info "PhoenixProtocol"
</li>
</ul>
(0029, 1020) is where the Phoenix Protocol is stored as a large text block alongside other fields.
The format looks like a long list of key = value pairs, something like:
<pre>tSequenceFileName = "*ep2d_bold" lRepetitions = 100 alTR[0] = 2000 alTE[0] = 30</pre>
<h3>Structure of the Phoenix Protocol</h3>
The protocol is essentially a plain-text configuration file embedded inside DICOM.
<ul>
<li>
Keys → Acquisition parameter names (tSequenceFileName, alTE[0], etc.)</li>
<li>
Values → Their corresponding settings (*ep2d_bold, 30, etc.)</li>
<li>
Arrays → Some parameters appear multiple times, e.g. alTE[0], alTE[1].</li>
</ul>
This makes it parsable with regular expressions.
<h3>Reading the Phoenix Protocol in Python</h3>
Here’s how you can extract and parse the Phoenix Protocol using Python.
You will need pydicom and nibabel for this task.
<pre>import re import pydicom from nibabel.nicom.csareader import get_csa_header # Path to a Siemens DICOM file dcm_path = "/path/to/your/siemens_csa_file.dcm" dicom_dataset = pydicom.dcmread(dcm_path) # Extract CSA header (series level) -- you can choose image level for other info-- csa_header = get_csa_header(dicom_dataset, "series") # Get the Phoenix Protocol text phoenix_protocol = csa_header["tags"]["MrPhoenixProtocol"]["items"][0] # Parse "key = value" pairs using regex matches = re.findall(r'(\S+)\s*=\s*([^\n\r]+)', phoenix_protocol) protocol_dict = {key: value.strip() for key, value in matches} # Example: print the sequence file name print(protocol_dict["tSequenceFileName"])</pre>
<h4>Example Output</h4>
If your scan used a Siemens EPI sequence, the output might look like:
<pre>'""%CustomerSeq%\\to_ep2d_VEPCASL""'</pre>
And your dictionary could contain hundreds of parameters like:
<pre>{ "tSequenceFileName": '""%CustomerSeq%\\to_ep2d_VEPCASL""', "lRepetitions": "100", "alTR[0]": "2000", "alTE[0]": "30" }</pre>
<h3>Why This Matters</h3>
<ul>
<li>
Research reproducibility → You can log and compare protocols across datasets.</li>
<li>
Quality assurance → Spot inconsistencies in TR/TE or slice setups.</li>
<li>
Automation → Build pipelines that read scan parameters directly from DICOMs instead of manual input.</li>
</ul>
<h3>Final Thoughts</h3>
The Siemens Phoenix Protocol is a treasure trove of metadata hidden inside DICOM files. With just a few lines of Python, you can parse and use it in your imaging workflows.
Whether you’re ensuring consistency across large datasets or just curious about what’s inside your MRI scans, this approach provides a powerful way to peek under the hood of Siemens DICOMs.
<h4>Useful Links:</h4>
<ul>
<li><a href="https://nipy.org/nibabel/dicom/siemens_csa.html">Siemens Format DICOM with CSA Header — Nibabel</a></li>
<li><a href="https://dicom-parser.readthedocs.io/en/latest/modules/dicom_parser.utils.siemens.csa.html">Dicom_Parser</a></li>
</ul>
<img src="https://medium.com/_/stat?event=post.clientViewed&amp;referrerSource=full_rss&amp;postId=d6c3dd9fd3f9" width="1" height="1" alt="">

How to Parse Siemens Healthineers Phoenix Protocol from CSA Headers

This project presents a comprehensive, end-to-end AI framework designed to address critical challenges in brain tumor segmentation from MRI scans. The work done in 2024 focuses on two major areas: mitigating the effects of motion artifacts and reducing the manual effort required for data annotation.
<h3>The Problem and Our Solution</h3>
Brain tumor segmentation from MRI scans is a complex task. Two major hurdles are:
<ol>
<li>
Motion artifacts: Patient movement during an MRI can distort images, making accurate segmentation difficult.</li>
<li>
Manual annotation: Manually outlining tumors in thousands of MRI slices is time-consuming, expensive, and prone to human error.</li>
</ol>
Our team developed a two-pronged approach to tackle these issues. First, we created an advanced deep-learning pipeline for motion artifact correction. This pipeline uses a novel technique to correct spatial misalignments caused by patient movement, significantly improving image quality. Second, we implemented an active learning framework to iteratively improve the segmentation model with minimal manual input.
<h3>AI Research and Technical Details</h3>
Our research work is the core of this project, focusing on the following key components:
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yU78fKBy_dV6yKjlGypFUA.png"><figcaption>Our Propsed Wavelet Stacked UNet</figcaption></figure><h4>Motion Artifact Correction</h4>
Our deep-learning pipeline for motion artifact correction is a significant technical achievement. It integrates co-registration as a crucial preprocessing step to align motion-free and motion-corrupted MRI volumes from the same patient. The pipeline utilizes Wavelet Stacked UNets with self-assisted priors and a combined loss function of DSSIM (structural similarity index) and Perceptual Loss.
<ul>
<li>
Wavelet Decomposition: This technique allows our model to effectively capture image contrast, detailed structures, and edges.</li>
<li>
Performance Improvements: By applying this pipeline to the BraTS2021 dataset, we saw impressive results. The SSIM (Structural Similarity Index Measure) score improved from 76.85% to 92.92%, and the PSNR (Peak Signal-to-Noise Ratio) score increased from 24.96 to 32.78. These metrics demonstrate a substantial improvement in the quality of the corrected images.</li>
<li>
Loss Function Optimization: We also proposed a new technique to optimize the weights and biases when combining multiple loss functions, a key aspect of training complex deep-learning models</li>
</ul>
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xUoVLc7O7yBnZy4afp1dFQ.png"><figcaption>Active Learning Framework — Real Life Scenario</figcaption></figure><h4>Active Learning for Annotation</h4>
To minimize the need for manual annotation, we implemented an active learning framework. This framework uses a pre-trained U-Net model, initially trained on the BraTS2021 adult brain MRI dataset. It uses uncertainty estimation to identify the most challenging or “uncertain” images for which the model needs human feedback.
<ul>
<li>
Iterative Refinement: By having an expert annotate only these uncertain images, the model can be refined iteratively, significantly reducing the overall manual effort while improving segmentation accuracy.</li>
<li>
Validation: This framework was successfully validated on the BraTS-PED dataset, demonstrating its effectiveness in a new context.</li>
</ul>
<h4>The End-to-End Imaging Platform</h4>
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HfY_4coVkhy-DnkoJSCW1w.jpeg"></figure>Beyond the core AI research, we developed a full-featured, cloud-based medical platform from scratch in collaboration with Ain Shams University Hospital (ASUH) — El Demerdash. This platform provides a user-friendly interface for clinicians to deploy our models and frameworks, facilitating seamless clinical integration. The system is built on a microservices architecture, ensuring scalability and modularity.
In developing our platform, we used the latest frontend and backend technologies in the medical software industry, adhering to international healthcare standards such as DICOM Standard for web-based medical imaging (DICOM-Web), for medical imaging interoperability, ISO 62304 for medical device software lifecycle processes, Health Insurance Portability and Accountability Act (HIPAA) for patient data protection, and Brain Imaging Data Structure (BIDS) for organizing and storing neuroimaging data.
<ul>
<li>We used React and TypeScript for a dynamic, type-safe user interface, with Cornerstone3D for 2D and 3D medical image visualization. Figma was used for designing low and high-fidelity wireframes.</li>
<li>The backend is powered with Django Rest Framework for managing data and creating RESTful APIs, PostgreSQL for the database due to its reliability and scalability, and RabbitMQ for asynchronous communication between services. Orthanc DICOM Server was integrated for the storage and retrieval of DICOM data. FastAPI was used for performance-critical endpoints.</li>
<li>
Python was extensively used for development, with Postman for API testing and documentation, ensuring endpoints were reliable and well-documented. Our development process included Docker for containerization, Git/GitHub for version control, and AWS for deployment.</li>
</ul>
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tJPSSoJPl-3lF59wXyjuWQ.png"><figcaption>System Design and Architecture for the platform</figcaption></figure>The proposed platform architecture is designed for handling deep learning for advanced image analysis while ensuring solid data management. It features a distributed system with loosely coupled services, consisting of two components: the Client Side, which is a web-based viewer that is accessible on various devices providing interaction via RESTful APIs; and the Cloud (Server Side), which hosts services for data processing and storage. The Gateway manages client requests, functioning as a load balancer, reverse proxy, and authentication manager. Orthanc and NIfTI Storage handle DICOM[1] and NIfTI data, respectively. The Message Broker enables asynchronous communication, enhancing system scalability and reliability. Inference servers host our proposed deep learning models for tasks like brain tumor segmentation, motion artifact correction, and MRI sequence synthesis. The Reporting service generates automatic reports enabling doctors to write reports for patient studies, accessible to medical professionals for review and editing. The Logging Service tracks system operations for monitoring, debugging, and auditing.
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gullTSGPeMgTw_kAsdfXjA.png"><figcaption>Data flow diagram</figcaption></figure>The system processes medical imaging data through a structured 3-tier application. Initially, imaging data from modalities like MRI and CT are stored in the PACS Server or NIfTI Storage, managing DICOM and neuroimaging data, respectively. The data undergoes processing, including metadata extraction and analysis via a deep learning pipeline. This pipeline enhances image quality and diagnostic value by removing motion artifacts with a correction model and synthesizing missing MRI sequences using an MRI sequence synthesis model. A pre-trained optimized U-Net model then creates high-accuracy tumor segmentation masks. The processed results, including segmentation masks, are accessible to medical professionals via our web application, which offers options for adjusting segmentation layers, adding annotations, and other data review and analysis functionalities.
<h3>Resutls</h3>
The proposed WAT block in our Stacked U-Nets with SAP model corrects motion artifacts on a new real-world dataset, has significantly improved the dataset metrics, increasing the SSIM score to 92.92% and the PSNR to 32.78. Furthermore, this enhancement has led to improved brain tumor segmentation results. While initial Active Learning iterations improved the model’s average dice score from 0.215 to 0.429 on the test set. Further exploration of hyperparameters and query strategies is crucial for optimizing segmentation accuracy.
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f2APiFGo9jjmMM9zJhdczQ.png"></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y-JXAJg7_O946B9O1JJRww.png"></figure><h3>Team</h3>
<h4>Members</h4>
<ul>
<li>
<a href="https://github.com/1brahimmohamed">Ibrahim M. Abdelazim</a> — Software Development</li>
<li>
<a href="https://github.com/MohamedAIsmail">Mohamed Ismail</a>— Software Development</li>
<li>
<a href="https://github.com/mahmoud1yaser">Mahmoud Salman</a> — Research &amp; AI</li>
<li>
<a href="https://github.com/mahamedhat">Maha Ezzelarab </a>— Research &amp; AI</li>
<li>
<a href="https://github.com/ahmedhassan187">Ahmed Hassan</a>— Research &amp; AI</li>
</ul>
<h4>Supervisors:</h4>
<ul>
<li>
<a href="https://scholar.google.co.kr/citations?user=y_8D7KEAAAAJ&amp;hl=en">Dr. Meena M. Makary</a> - Associate Professor, Systems &amp; Biomedical, Cairo University - Research Scientist | Radiology, Harvard Medical School</li>
<li>
<a href="https://scholar.google.com/citations?user=7nV1crUAAAAJ&amp;hl=en">Dr. Mohammed Al-masni</a> - Assistant Professor, Department of Artificial Intelligence, Sejong University, South Korea</li>
</ul>
<h3>Acknolwedgement</h3>
We sincerely thank Ain Shams University Hospital for their support. Special thanks to Dr. Khaled Abouelfotouh, Dr. Shaimaa Abdelsattar, and Dr. Marwa Mohamed for their valuable assistance and collaboration.
<figure><img alt="" src="https://cdn-images-1.medium.com/max/170/1*ICifzY7AVUY4qvCOw-NE8w.png"></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*FNq6ZiiRgU2hu0mCRp-H2A.png"></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/983/1*RuU2b5S6vuBVk8DoCRV4Vw.png"></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a-RD82ah6DQqN3fAjWo4qg.png"></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/337/1*JEoLkjBKj4vH7720jnycYA.png"></figure><img src="https://medium.com/_/stat?event=post.clientViewed&amp;referrerSource=full_rss&amp;postId=6ab4eedc0e83" width="1" height="1" alt="">

Graduation Project -Brain Tumor Segmentation: An AI-Based Framework

<figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*2oZsCg2OX6IpH-8eI2xqSA.png"></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wc0X0C_EYpURiISKo8KYHA.png"></figure>This summer, I had the opportunity to participate in Google Summer of Code (GSoC) 2025 with the Open Science Initiative for Perfusion Imaging (OSIPI). My project focused on enhancing the ASL Parameters Report Generator, a tool that extracts and reports key perfusion imaging parameters from DICOM and BIDS datasets. Perfusion imaging, particularly Arterial Spin Labeling (ASL), plays a crucial role in assessing cerebral blood flow in clinical and research settings. However, reporting ASL parameters in a standardized, reproducible, and user-friendly way has been a challenge. My work aimed to address this by refactoring the existing prototype into a reusable Python package and building a modern web application around it.
<h3>Project Goals</h3>
The goal of my project under OSIPI was to improve and extend the ASL Parameters Report Generator. The previous year’s work was a web application prototype with logic tightly coupled to the UI.
My project focused on:
<ul>
<li>Extracting the core functionality into a standalone Python package to improve modularity, maintainability, and reusability.</li>
<li>Enhance the Code quality and readability</li>
<li>Building a new FastAPI backend and Next.js frontend that consumes the Python package and provides an improved user experience.</li>
<li>Adding support for parameter extraction from DICOM images in addition to BIDS (previously, the tool only supported BIDS).</li>
<li>Including missing parameters and enabling PDF export of the generated report.</li>
<li>Creating user and developer documentation and introducing Dockerization to simplify setup and reproducibility.</li>
</ul>
<h3>What I Did</h3>
<ul><li>
Core Refactoring into Python Package - Separated application logic from the web interface - Designed a reusable Python package with well-documented functions for ASL parameter extraction. - Published the package with a clear structure for future contributors</li></ul>
<figure><img alt="" src="https://cdn-images-1.medium.com/max/851/1*fnCYULrK997Y2pD1aje58g.png"><figcaption>Python Package Architecture</figcaption></figure><ul>
<li>
DICOM-to-BIDS Conversion &amp; Parameter Extraction - Implemented new logic for extracting parameters directly from DICOM images. - Converted the dictionary of ASL parameters into executable Python code. - Unified workflows for BIDS and DICOM datasets to ensure consistency.</li>
<li>
New Web Application (FastAPI + Next.js) - Built a FastAPI backend that integrates with the Python package to serve parameter extraction and reporting as APIs. - Developed a Next.js frontend with an intuitive interface for uploading data, viewing extracted parameters, and exporting reports. - Implemented light/dark mode and improved layout for readability.</li>
</ul>
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BjWxUPR-wxhGccQ-K6Zbbw.png"><figcaption>Whole Application Flow and Technologies</figcaption></figure><ul>
<li>
Feature Enhancements - Added PDF export of reports (previously only JSON/HTML - Provide missing parameters in the generated report. - Improved error handling and validation for user uploads.</li>
<li>
Documentation &amp; Deployment - Created a documentation website explaining installation, usage, and developer guidelines. - Enhanced Docker workflows to simplify local setup and ensure reproducibility. - Wrote clear instructions for contributors and users.</li>
</ul>
<h3>3. Current State of the Project</h3>
<ul>
<li>The Python package is complete, modular, and ready for reuse in other projects.</li>
<li>The web application is functional and provides a polished UI/UX.</li>
<li>Reports can be generated from both DICOM and BIDS datasets.</li>
<li>Export to PDF and inclusion of missing parameters are implemented.</li>
<li>Documentation is available and tested.</li>
<li>The project is containerized with Docker for quick deployment.</li>
</ul>
<h3>What’s Left to Do</h3>
<ul>
<li>Further validation on larger and more diverse DICOM/BIDS datasets.</li>
<li>Extend reporting to cover other ASL sequences and other Perfusion Modalities like DCE, and ensure compliance with OSIPI standards.</li>
<li>Add automated tests for the Python package to strengthen reliability and more robust CI/CD pipeline.</li>
<li>Possible integration with other OSIPI tools for end-to-end workflows.</li>
</ul>
<h3>Code Contributions</h3>
<ul>
<li>
Main Repository: <a href="https://github.com/1brahimmohamed/ASL-Parameter-Generator">https://github.com/1brahimmohamed/ASL-Parameter-Generator</a>
</li>
<li>
Python Package: <a href="https://pypi.org/project/pyaslreport/">https://pypi.org/project/pyaslreport/</a>
</li>
<li>
Web Application (FastAPI + Next.js): <a href="https://asl-parameter-generator.vercel.app/">https://asl-parameter-generator.vercel.app/</a>
</li>
<li>
Documentation Website: <a href="https://docs.page/1brahimmohamed/ASL-Parameter-Generator">https://docs.page/1brahimmohamed/ASL-Parameter-Generator</a>
</li>
</ul>
<h3>Challenges &amp; Learnings</h3>
<ul>
<li>Migrating from a monolithic web app to a reusable Python package required redesigning the architecture with modularity and clean APIs in mind.</li>
<li>Translating a written dictionary into working extraction code was challenging, but it gave me a deeper understanding of ASL parameters</li>
<li>Although I had prior experience, I discovered better practices for multi-service container setups and reproducibility.</li>
<li>Creating a dedicated documentation website taught me the importance of making work accessible for both users and future contributors.</li>
<li>Working with mentors and collaborators improved my communication, planning, and open-source contribution skills.</li>
</ul>
<h3>Acknowledgments</h3>
I would like to thank my mentors ( Dr. Jan Petr (<a href="https://scholar.google.com/citations?user=qS-vflwAAAAJ&amp;hl=en">Google Scholar)</a>, Dr. David Thomas <a href="https://profiles.ucl.ac.uk/6699-david-thomas">(Website)</a>, and Hanliang Xu <a href="https://github.com/Hanliang-Xu">(Github)</a> ) for their continuous support, responsiveness, and guidance. which made collaboration smooth and enjoyable. Their feedback and encouragement were invaluable throughout the program.
<img src="https://medium.com/_/stat?event=post.clientViewed&amp;referrerSource=full_rss&amp;postId=2f95fab822ce" width="1" height="1" alt="">

Informative

Recent Articles