Reinventing the X-ray Wheel — Together: How CONRAD is Transforming Cone-Beam Imaging
A Beginning of New Era in Medical Imaging
In the ever-evolving world of medical imaging, innovation is often hidden beneath the surface — encoded deep within the lines of software that bring our diagnostic technologies to life. In 2013, a groundbreaking paper quietly revolutionized this landscape. Published in Medical Physics, it introduced CONRAD, a comprehensive, open-source software framework for cone-beam imaging in radiology. The work, led by Andreas Maier and a transatlantic team of engineers and clinicians from Stanford University and Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), is far more than a technical achievement — it’s a bold declaration of open science and collaborative progress in medical technology.
But why should anyone outside a radiology lab care about software code and projection matrices? Because the implications are massive: improving imaging accuracy, enabling cross-lab collaboration, reducing redundant work, and, ultimately, enhancing patient care. CONRAD’s very existence invites us to rethink how science is done — and shared.
Why This Paper Matters
Before CONRAD, the world of cone-beam computed tomography (CBCT) software was fragmented. Research groups often built their own closed systems, leading to redundancy, lack of comparability, and isolation. Every lab implemented the same algorithms slightly differently, which meant that cross-validation of new methods was fraught with uncertainty. Even sharing software was a minefield of licensing, undocumented code, and intellectual property concerns.
Enter CONRAD — short for CONe-beam in RADiology — a software framework that serves as both a toolbox and a bridge. It was built to unify simulation and reconstruction processes in CBCT, while also promoting transparency, academic credit, and open collaboration. With its launch, for the first time, the medical imaging community had a versatile, extensible, and scientifically rigorous open-source platform that could simulate real-world x-ray systems and reconstruct high-quality images using cutting-edge algorithms.
A Thoughtful Blueprint for Collaborative Imaging
The CONRAD framework isn’t just a bunch of code. It’s an ecosystem, designed with principles that elevate it from a technical convenience to a scientific infrastructure. Its architecture reflects a deep understanding of both the technological and sociological challenges in the field.
At its core, CONRAD is written in Java, making it inherently platform-independent and accessible on both Windows and Linux. It integrates hardware acceleration through OpenCL, enabling high-performance computing on GPUs. And crucially, it uses a streaming pipeline architecture, where each image-processing step is treated as a modular, interchangeable component — much like LEGO blocks. This makes experimentation and optimization both intuitive and scalable.
Even more impressively, every module within CONRAD includes proper scientific citations. This means when you run an algorithm, the software tells you exactly who developed it and where it was published. It’s software that teaches, credits, and encourages proper attribution — an elegant merger of code and academic ethics.
Simulation from the Ground Up
One of CONRAD’s superpowers lies in its simulation capabilities. Generating realistic x-ray projection images requires modeling geometry, motion, and physics with extreme precision. CONRAD handles all of this.
First, it allows users to model anatomical phantoms using simple shapes or advanced descriptors like NURBS and surface splines. It even supports FORBILD phantom definitions — a gold standard in phantom modeling. Users can define motion models through a system of interfaces that support periodic and nonlinear motion via time warpers, a concept drawn from mathematical signal transformations. These models allow for accurate 4D simulations of moving organs like lungs or hearts.
Next, CONRAD handles physical absorption using monochromatic or polychromatic x-ray models. It integrates with the NIST XCOM database, enabling accurate computation of x-ray attenuation across materials and energy spectra. Simulated photon statistics? No problem — Poisson noise can be layered in to simulate realistic detector readings.
Finally, using ray casting, it computes projection data pixel-by-pixel, calculating every interaction along a ray’s path through a virtual patient. This isn’t just a rendering trick; it’s a physics-informed reconstruction of what the x-ray detector would actually measure in real life.
Reconstruction with Real-World Power
On the reconstruction side, CONRAD doesn’t shy away from real-world complexity. It offers an impressive suite of correction algorithms — like beam hardening, scatter correction, and cosine weighting — to address the nonlinear artifacts and inaccuracies that arise in clinical scans.
It supports advanced noise reduction methods, including anisotropic adaptive filtering guided by structure tensors — essentially using image structure to smartly smooth out noise without losing important detail. Reconstruction algorithms span the spectrum from Filtered Back Projection (FBP) to Algebraic Reconstruction Techniques (ART), with multi-core and GPU-parallelized versions for performance.
Even partial data and short scan scenarios are addressed. CONRAD integrates redundancy weighting methods from the likes of Parker and Noo, ensuring robust image quality even when acquisition conditions aren’t ideal.
Demonstrating Excellence: Real Experiments, Real Results
A framework is only as good as its results, and CONRAD delivers. The team successfully reconstructed images from both a tabletop CT system and a clinical C-arm scanner.
In one experiment, a Catphan 500 phantom was imaged with a Varian detector, producing 360 projections with 1° angular increments. The resulting 3D reconstruction (512×512×50 voxels) had a voxel size of 0.4×0.4×0.5 mm and revealed excellent fidelity.
In a second test, a clinical Siemens Artis zeego system scanned a pig specimen, acquiring 661 projections. The reconstruction volume was a hefty 51²³ voxels at 0.5 mm resolution. Despite the complexity, CONRAD handled it with elegance, delivering clinically relevant images.
To top it off, CONRAD’s OpenCL back-projector achieved blazing-fast runtimes: 2.5 seconds for a 256³ voxel volume and 12.4 seconds for a 512³ volume — outperforming even CUDA-based implementations in the RabbitCT benchmark.
A Coalition of Visionaries
This endeavor was no solo mission. It took the coordinated effort of researchers from multiple institutions:
Andreas Maier, Jang-Hwan Choi, Christian Riess, Andreas Keil, and Rebecca Fahrig — Department of Radiology, Stanford University
Hannes G. Hofmann, Martin Berger, Peter Fischer, Chris Schwemmer, Haibo Wu, Kerstin Müller, and Joachim Hornegger — Pattern Recognition Lab and Erlangen Graduate School in Advanced Optical Technologies (SAOT), Friedrich-Alexander University Erlangen-Nürnberg
Funded by the NIH, the Lucas Foundation, and Germany’s DFG Excellence Initiative, this project exemplifies the kind of large-scale, cross-border cooperation that drives scientific progress.
The Bigger Picture: Why CONRAD Matters
The brilliance of CONRAD is not just technical — it’s cultural. It redefines how medical imaging software should be built, shared, and credited. By embedding scientific recognition directly into its core, it enables a new form of collaboration. Labs don’t have to reinvent the wheel. Instead, they can build together on a shared, solid foundation.
For clinicians, this means faster integration of new algorithms into real-world applications. For patients, it could mean quicker, more accurate diagnoses. For researchers, it’s a pathway to reproducibility and impact.
And it’s all open-source — available for download at https://github.com/akmaier/CONRAD. You can run it, modify it, contribute to it, and cite it. The project includes tutorials, nightly builds, and full documentation. Even MATLAB integration is supported.
Looking Forward: From Phantom to Clinic
As imaging science shifts toward personalized, motion-aware, and AI-supported workflows, frameworks like CONRAD become even more vital. Already supporting 4D reconstruction and GPU acceleration, it lays the groundwork for integrating machine learning, real-time processing, and patient-specific modeling.
Whether it’s developing radiation-reducing protocols, simulating surgical scenarios, or training new generations of imaging scientists, CONRAD is the backstage workhorse quietly powering the future of radiology.
This paper, and the framework it introduces, offers something rare: a software toolkit built not just for today’s researchers, but for the collaborative, open, and interdisciplinary science we need tomorrow.
And that — quietly, decisively — is how you reinvent the x-ray wheel. Together.
This blog post is based on this 2013 Medical Physics Paper.
If you liked this blog post, I recommend having a look at our free deep learning resources or my YouTube Channel.
Text and images of this article are licensed under Creative Commons License 4.0 Attribution. Feel free to reuse and share any part of this work.