The retrospective, multicentre study, led by Mani Arsalan of the University Hospital of the Goethe University in Frankfurt and colleagues across German heart centres, enrolled 247 patients with symptomatic severe aortic stenosis who underwent TAVI at two high-volume centres. Each patient’s pre-procedural CT dataset was analysed twice, independently. The first analysis used 3mensio Structural Heart software, version 10.1 from Pie Medical Imaging, the widely adopted semi-automated standard in which experienced operators manually adjust the lumen centreline, define the annulus plane by marking the nadirs of the aortic leaflets, and then measure the average diameter, perimeter and area of the annulus along with distances to the coronary arteries. The second analysis was performed by heart.ai, a cloud-based platform from Laralab GmbH in Munich that applies deep learning algorithms to segment the aortic annulus and surrounding cardiac structures with no user intervention whatsoever.

The technical foundations of the AI platform are substantial. Its convolutional neural network models were trained on 811 CT datasets comprising more than 250,000 image slices, drawn from routine clinical practice across multiple institutions, different geographical locations, and a wide range of CT scanners and image qualities, including high-resolution and non-contrasted acquisitions. Roughly 80 to 90 percent of the data were used for model training and 10 to 20 percent for internal validation, with a strictly separated external test set reserved for final evaluation. The architecture is deliberately modular: one three-dimensional convolutional network specializes in segmenting the aortic cusps, identifying the nadirs that define the annular plane, while another dedicated network segments the aortic root, whose outer contour is then used to derive the annulus itself. Characteristic heart planes for multiplanar reconstruction are calculated from the resulting three-dimensional models, and custom algorithms automatically extract the measurements clinicians need. Datasets with slice thickness above 3.0 millimetres or significant numbers of missing slices are automatically rejected.

When the two methods were compared head to head, the agreement was striking. The mean aortic annulus diameter measured 24.5 plus or minus 2.3 millimetres with 3mensio and 24.4 plus or minus 2.4 millimetres with the AI platform, yielding a mean absolute error of just 0.6 millimetres, a mean absolute percentage error of 2.6 percent, and an intraclass correlation coefficient of 0.968. Annulus perimeter showed similarly tight concordance, at 76.9 versus 74.6 millimetres with a mean absolute error of 2.0 millimetres, while annulus area came in at 458.4 versus 440.9 square millimetres, a mean absolute error of 21.2 square millimetres. All three annular parameters achieved intraclass correlation coefficients above 0.95, a threshold the researchers classify as excellent agreement. Bland-Altman plots confirmed that the differences between methods were small and consistent across the measurement range.

Coronary ostial distances, which determine how close the implanted valve will sit to the arteries supplying the heart muscle, also showed high concordance, though with somewhat weaker correlations than the annular measurements. The mean distance to the right coronary artery was 17.5 millimetres by the conventional method versus 16.8 millimetres by AI, with an intraclass correlation coefficient of 0.892; the corresponding left main coronary artery figures were 14.2 versus 13.5 millimetres, with a coefficient of 0.886. The researchers note that this relative softening reflects the greater anatomical variability and segmentation challenges these structures pose, and that even here the mean absolute errors remained small enough that they may not be clinically meaningful.

The most consequential question, however, was not whether the machines could measure like humans, but whether they would choose like them. When the researchers retrospectively simulated prosthesis size selection based solely on each method’s measurements, applying the manufacturers’ instructions for use, AI-derived measurements would have led to a different valve size than the one actually implanted in 21 percent of patients. The semi-automated 3mensio-based simulation would have produced a different choice in 14 percent. Both figures fall within the 10 to 20 percent range of inter-software variation reported in other recent comparative studies, and most discordances involved neighbouring valve sizes rather than dramatic mismatches. Crucially, the authors emphasize, CT measurements are never used in isolation in clinical practice; the final valve choice integrates the full anatomical and clinical picture, including the left ventricular outflow tract, the sinus of Valsalva, and the amount and distribution of calcification.

The study’s findings align with a rapidly growing literature. Santalo-Corcoy and colleagues validated the TAVI-PREP deep learning tool in 200 patients and demonstrated expert-level concordance while cutting analysis time to roughly two minutes, compared with more than fifteen minutes for conventional manual workflows. Wang and colleagues, in a large multicentre study, reported agreement coefficients above 0.96 between a fully automated algorithm and manual measurements, closely mirroring the present results. There is also a compelling physical argument for machine precision: CT data span more than 1,000 Hounsfield units, while the human eye can distinguish only about 30 shades of grey. Even with windowing, subtle density differences remain invisible to human readers but are readily detectable by AI, allowing it to differentiate tissues with nearly identical densities, such as non-contrasted blood versus myocardium, or calcification versus contrast agent.

The practical appeal of full automation extends well beyond raw accuracy. In high-volume TAVI centres, dramatic reductions in analysis time, lower operator dependency and improved reproducibility translate into real resource savings and expanded workforce capacity, particularly for urgent or complex cases. Automated platforms could also serve as a digital second opinion, flagging borderline anatomies for closer expert scrutiny. Recent work suggests the scope will only widen: Inomata and colleagues showed that a deep learning model can accurately quantify aortic valve calcification on contrast-enhanced CT, consistent with manual Agatston scoring, opening the door to fully integrated automated calcification assessment within TAVI planning pipelines. Yet the authors are careful to temper enthusiasm. Complex anatomies, heavy calcification, a horizontal aorta and bicuspid valves could all exacerbate measurement differences, and in this study every automated measurement still required confirmation by a human reader before final acceptance.

The study has limitations that the researchers acknowledge. As a retrospective analysis focused on agreement between methods, it did not assess analysis time or inter- and intraobserver variability, and it cannot establish whether AI-driven planning would actually improve procedural complications or long-term valve function. That question awaits prospective, outcome-oriented trials. The authors also point toward a future in which AI integrates with biomechanical modelling, radiomics and predictive analytics, pushing TAVI planning beyond static measurement into genuine outcome forecasting, with multicentre studies needed to validate performance across diverse populations, including patients with bicuspid valves and other challenging anatomies. For now, the message is one of measured optimism: fully automated AI-based CT analysis has proven it can match the precision of the established semi-automated gold standard, offering speed, standardization and objectivity, while the judgment of the human expert remains, for the foreseeable future, the final arbiter of which valve goes into which heart.

Subject of Research: Fully automated AI-based CT analysis compared with semi-automated software for pre-procedural TAVI planning and valve sizing

Article Title: Impact of fully-automated AI based CT-analysis on pre-procedural TAVI planning

Article References: Arsalan, M., Duske, T., Schneider, H., Tamm, A. R., Seppelt, P. C., Geyer, M., Piayda, K., von Bardeleben, R. S., Martin, S., Leistner, D., Hell, M., Walther, T., & Kreidel, F. (2026). Impact of fully-automated AI based CT-analysis on pre-procedural TAVI planning. Clinical Research in Cardiology. https://doi.org/10.1007/s00392-026-03017-y

Image Credits: AI Generated

DOI: 10.1007/s00392-026-03017-y

Keywords: TAVI, aortic stenosis, artificial intelligence, deep learning, CT analysis, aortic annulus, prosthesis sizing, cardiology, medical imaging, convolutional neural networks, valve replacement, clinical research