Conference Agenda

A major challenge in advancing functional organ assessment is the limited understanding of vascular development and morphology. To address this, we present SynGROW (Synthetic Generation of Rigorously and Rapidly Optimized Volumetric Vasculature) – our open-source computational framework for generating and analyzing synthetic vascular trees using rigorous global optimization techniques [1]. Our objective function minimizes both the energy required to maintain the vascular structure and the energy required to pump the blood through it, while incorporating patient-specific information from imaging data. Our physiological model does not impose Murray’s law of optimal branching explicitly; instead, we show that it naturally emerges from the optimization process [2]. Our model also incorporates variable blood viscosity at different levels of the vessel hierarchy due to the Fåhræus-Lindqvist effect.

SynGROW integrates global geometry and topology optimization into a unified algorithm that solves a nonlinear programming (NLP) problem with super-linear efficiency. Unlike existing methods limited to single-tree generation in convex domains, SynGROW contains various algorithmic extensions that enable the generation of multiple non-intersecting vascular tree structures within complex nonconvex anatomical domains [3].

In benchmark tests on brain tissue, SynGROW outperforms state-of-the-art methods based on local algorithms, computing results an order of magnitude faster while significantly reducing the objective function. Its results closely match experimental liver vasculature data, accurately capturing physiological features such as tortuosity and parallel branching. Beyond benchmarks, we demonstrate its integration within functional imaging and multiphysics modeling workflows. In the brain, it predicts spatial patterns of penetrating arteries and veins aligned with observed microcirculation [4]. In the liver, it serves as a foundation for modeling hepatic hyperperfusion in the context of liver regeneration [5].

References

[1] E. Jessen, M.C. Steinbach, C. Debbaut and D. Schillinger. Rigorous mathematical optimization of synthetic hepatic vascular trees. J. R. Soc. Interface 19:20220087, 2022.

[2] E. Jessen, M.C. Steinbach, C. Debbaut and D. Schillinger. Branching exponents of synthetic vascular trees. IEEE Trans. Biomed. Eng. 71:1345-1354, 2024.

[3] E. Jessen, M.C. Steinbach, and D. Schillinger. Optimizing non-Intersecting synthetic vascular trees in nonconvex organs. IEEE Trans. Biomed. Eng. (in press).

[4] L.B. Glandorf, E. Jessen, et al. Deciphering morpho-functional stroke repercussions on the cortical microvasculature with Bessel beam OCM. In: Proc. Optical Coherence Tomography and Coherence Domain Optical Methods in Biomedicine XXIX, SPIE, 2025.

[5] A. Ebrahem, J. Hohl, E. Jessen, M.F.P. ten Eikelder and D. Schillinger. A multiscale-multiphysics framework for modeling organ-scale liver regrowth. J. Mech. Phys. Solids 200:106113, 2025.

The diagnosis of cardiac ischemia in acute conditions is of utmost importance in the clinical practice since several therapeutic decisions could follow. In the first part of this talk, we introduce a new cardiac perfusion model [1], accounting for the coronary compliance and for the diastolic flow of microcirculation, for the non-invasive prediction of blood flow maps at the myocardial level. We also discuss the challenging issue of its calibration to describe personalized conditions of patients. Finally we show an application of this model to a wide cohort of patient coming from Centro Cardiologico Monzino in Milan.

In the second part of the talk, we address the issue of assessing the arrhythmogenic risk, in terms of development of ventricular tachicardia, in presence of acute ischemic regions. To do this, we start from the maps discussed in the first part of the talk and we present a computational model [2] based on the local modification of the ionic currents to account for specific processes (such as hyperkalemia and hypoxia) which are consequences of the presence of acute ischemia. We present results of numerical experiments aiming at evaluating, among the others, the effect of the ischemic heterogeneity on the arrhythmogenic risk.

All the numerical experiments were performed with the Finite Elements library LifeX, developed at MOX, Dipartimento di Matematica, in collaboration with LaBS, Dipartimento di Chimica, Materiali e Ingegneria Chimica, both at Politecnico di Milano

REFERENCES

[1] Montino Pelagi G., Regazzoni F., Huyghe J.M., Baggiano A., Ali' M., Bertoluzza S., Valbusa G., Pontone G., Vergara C., Modeling cardiac microcirculation for the simulation of coronary flow and 3D myocardial perfusion. Biomech. Model. Mechanobiol., 23, 1863-1888, 2024.

[2] Corda A., Pagani S., Vergara C., Influence of acute myocardial ischemia on arrhythmogenesis: a computational study, medRXiv, DOI:10.1101/2024.11.20.24317476, 2024

Introduction: Aortic stenosis (AS) is the most prevalent primary valve disease requiring intervention. Transcatheter aortic valve implantation (TAVI) has become the preferred treatment for elderly patients at high surgical risk. However, procedural complications remain a major concern. Among these, conduction abnormalities—affecting 2% to 34% of cases—are primarily caused by excessive mechanical stress exerted by the bioprosthesis on the membranous septum, often leading to the need for permanent pacemaker (PPM) implantation. Additionally, paravalvular leakage (PVL), observed in 7% to 40% of TAVI patients, primarily results from an inadequate sealing between the prosthesis and the aortic annulus.

To address these challenges, in-silico technologies are increasingly used to simulate TAVI procedures. However, existing models have limitations in predicting complications, and their high processing time hinder their clinical applicability. For this reason, this study aims to develop a machine learning (ML) model trained on clinical data and to compare its predictions with the outcomes of patient-specific TAVI simulations. By integrating these approaches, we seek to create a comprehensive predictive profile for each patient, enabling the prognosis of conduction disturbances and PVL. Ultimately, this could potentially replace complex numerical simulations, facilitating more efficient treatment planning.

Methods: A total of 1,897 patients undergoing TAVI were enrolled and their clinical, laboratory, echocardiographic, and procedural data were collected. After conducting exploratory data analysis and feature engineering, dimensionality reduction was performed with UMAP. A Topological Data Analysis (TDA) Kaplan Mapper was then applied to identify patient phenotypes more susceptible to complications. To address class imbalance, undersampling of the majority class was performed, ensuring a 1:1 ratio. After that, various machine learning classifiers were trained and evaluated for their ability to predict conduction abnormalities and PVL.

Concurrently, an external cohort of 50 patients was recruited for in-silico modeling. TAVI simulations were conducted, incorporating patient-specific anatomy, the implanted bioprosthesis, and arterial pre-stress. TAVI digital twins were then validated against post-procedural clinical data. Finally, the external patient cohort will serve as an independent validation set for the ML model. Predictive outcomes will be analyzed alongside patient-specific simulations to assess the model’s limitations and improve its interpretability.

Results: TDA revealed effective clustering of patients who developed moderate or more severe PVL and required PPM implantation. The ML models demonstrated strong performance, with an accuracy of 0.79 and an F1 score of 0.80 for predicting PVL, and an accuracy of 0.76 and an F1 score of 0.78 for predicting PPM implantation. The classifiers that yielded the best performance were XGBoost and Random Forest, respectively. In terms of in-silico modeling, TAVI digital twins successfully replicated valve positioning within the aortic root, as validated by angiographic images and post-operative CT stent segmentation, with errors lower than 5% in terms of orifice area. Furthermore, diastolic CFD simulations were successfully completed.

Conclusions: Our findings are highly promising, demonstrating effective AI-based prediction of complications and accurate in-silico reproduction of the TAVI procedure. Refining ML models and integrating them with patient-specific simulations could significantly enhance clinical practice by providing a robust, rapid, and automated tool for pre-operative planning.