Automated extraction of novel biomarkers from routine imaging examinations, e.g., opportunistic CT, is a promising opportunity to extend current screening possibilities and to better guide individualized therapies. In theory, opportunistic CT can be used to obtain quantitative biomarker data from any tissue included in a routine examination.
Over the past decade, AI has been used to develop many different automated pipelines for such extractions. These include volumetric fat measurements in patients with diabetes or metabolic syndrome, as well as the extraction of bone mineral density (BMD) in osteoporotic patients. To unlock the full potential of biomarker extraction, however, any automated pipeline must be robust, reproducible, and precise. Assessing precision errors is of enormous importance, especially when treatment decisions might be drawn from those measurements in large populations.
In the article presented here, our data shows that the application of contrast agent itself and the contrast phase during image acquisition represents a significant bias for BMD measurements that should not be neglected, especially when using an automated pipeline. Fortunately, this bias can be easily eliminated by linear regression. Although the universal and globally accepted DICOM imaging format allows for the storage of information about contrast administration, in our experience, documentation of the contrast phase is recorded sparsely and inconsistently in clinical practice, limiting easy extraction by automated pipelines.
To overcome this challenge, we developed and successfully tested an artificial neural network (ANN) that reliably detects the contrast phase on a given CT scan. By doing so, we were able to minimize the contrast-related error in BMD measurements.
The ANN is accessible on GitHub (https://github.com/ferchonavarro/anatomy_guided_contrast_ct) and has already been integrated into our freely available web tool Anduin (https://anduin.bonescreen.de), which is available to professionals and individuals for fully automated spine processing, fracture detection, and BMD extraction.
- A 2D DenseNet with anatomy-guided slice selection achieved an F1 score of 0.98 and an accuracy of 98.3% in the test set. In a public dataset, an F1 score of 0.93 and an accuracy of 94.2% were obtained.
- Automated adjustment for contrast injection improved the accuracy of lumbar bone mineral density measurements (RMSE 18.7 mg/ml vs. 9.45 mg/ml respectively, in the portal-venous phase).
- An artificial neural network can reliably reveal the presence and phase of iodinated contrast agent in multidetector CT scans (https://github.com/ferchonavarro/anatomy_guided_contrast_c). This allows minimizing the contrast-induced error in opportunistic bone mineral density measurements.
Authors: Sebastian Rühling, Fernando Navarro, Anjany Sekuboyina, Malek El Husseini, Thomas Baum, Bjoern Menze, Rickmer Braren, Claus Zimmer & Jan S. Kirschke