Julian A. Westphal, Philipp Kaess, Lea Mantz, Rashid A. Barnawi, Sami Saidi, Matthias P. Fabritius, Alexander Ziller, Georg Kaissis, Daniel Rueckert, Florian J. Fintelmann
5 min
Large-scale CT datasets are essential for modern medical research, yet they are often plagued by unreliable or missing metadata regarding whether intravenous contrast was administered and which body part was imaged. Manually verifying this information is labor-intensive and prone to human error. The researchers developed FALCON (Fully Automated Labeling of CT anatomy and intravenous CONtrast) to provide an open-access, automated solution for these classification tasks.
The researchers trained four separate ResNet9-based convolutional neural networks to classify body parts (head and neck, chest, or abdomen and pelvis) and detect the presence of intravenous contrast. The model uses a pipeline that selects a single central axial slice from a 3D volumetric CT scan to perform these predictions. The team validated the model using six independent, multicenter datasets comprising 3,138 CT scans from 3,126 patients, ensuring the model was tested on diverse, unseen data acquired over nearly three decades.
FALCON demonstrated exceptional performance, achieving F1 scores for contrast detection of 99.4% for head and neck, 98.3% for chest, and 98.1% for abdomen and pelvis scans in external testing. Body part classification was even more robust, achieving a perfect 100% F1 score on unseen data. Furthermore, the model is significantly faster than human experts; for instance, it requires only 1.3 seconds to annotate a head and neck CT scan, compared to the 21 seconds typically required by a human expert. The model is available as an open-access tool with a graphical user interface, making it easily deployable in research workflows.
By automating the curation of large-scale imaging datasets, FALCON removes a significant bottleneck in medical research. Its high accuracy and speed allow researchers to quickly organize and validate large archives of CT scans, facilitating more reliable downstream tasks like radiomics and body composition analysis. Because it is open-source and does not require specialized hardware, it is highly accessible for clinical and academic institutions.
Sam: Okay, so it's a map of numbers. Does the contrast dye change those numbers in a predictable way?
Alex: It does. When a special fluid — intravenous contrast — is injected into a patient's veins before the scan, it makes blood vessels and certain organs appear significantly brighter on that density map. The model is trained to recognise that specific brightness shift as a reliable signal that contrast was used.
Sam: And I gather it doesn't just look for contrast everywhere at once — it checks the body part first?
Alex: Exactly, and that's what makes the architecture clever. The system works in stages. First, a classifier identifies which part of the body is being scanned — chest, abdomen, and so on. Then, knowing the region, it applies the appropriate rules to look for the contrast signal. A bright patch that signals contrast in the chest might look completely different from one in the abdomen, so knowing the context matters.
Sam: What kind of AI model is doing all this pattern recognition?
Alex: It uses an architecture called ResNet9. Think of it as a series of filters stacked on top of each other, each one looking for progressively more complex patterns. The early filters detect simple things like edges and boundaries. The deeper filters learn to recognise more complex shapes — the outline of a kidney, the curve of the aorta. Together, they build up a kind of visual understanding of the image.
Sam: And how did they verify that it actually works? Did they test it against data it had never seen before?
Alex: They did, which is the critical test. They validated the model using scans from multiple different hospitals — institutions that had no involvement in training the system. The model achieved high accuracy on those external datasets, and in some cases its labels were actually more consistent than the original metadata already stored in the files.
Sam: That last point is telling. It's not just matching human performance — it's correcting errors that humans had already introduced into the records.
Alex: That's a fair reading of it. The researchers are careful not to overstate this, but it does suggest the original DICOM labels in large archives are less reliable than many researchers assume. The broader point of the paper is that before you build any study on top of a large imaging dataset, you need a way to verify what you actually have.
Sam: And by releasing FALCON as an open-access tool, other research teams can run it on their own archives without having to build something like this from scratch.
Alex: That's the intention. The researchers hope it lowers the barrier for anyone working with large CT datasets — so that cleaning up the data becomes a routine first step rather than a significant project in itself. Ultimately, more reliable labels should lead to more reliable research findings. Thanks for listening to ResearchPod.