This study aims to enhance and streamline intracardiac 4D Flow magnetic resonance imaging (MRI) processing by increasing automation for the quantitative and systematic assessment of left ventricular (LV) dysfunction. The study is designed to achieve the following three objectives.

The primary objective is to develop a convolutional neural network (CNN)-based deep learning model for the automatic segmentation of the LV endocardial contour throughout the cardiac cycle using intracavitary MRI data. To support model training, a dataset of LV endocardial wall segmentations will be generated from balanced steady-state free precession (bSSFP) images. A purpose-built retrospective MRI database of bSSFP images will be retrieved to accelerate training set creation.

The secondary objective is to develop a numerical framework for non-invasive MRI-based pressure-volume (PV) loop reconstruction and calculation of simplified hemodynamic force descriptors (HDFs). A prospective cohort of patients with severe aortic stenosis undergoing transcatheter aortic valve replacement (TAVR) will be enrolled. Pre-procedural non-contrast 4D Flow MRI will be acquired, and non-invasive MRI-derived PV loops will be quantitatively compared with invasive catheter-based PV loop measurements. In addition, simplified HDFs will be compared with 4D Flow-derived HDFs to assess their agreement and their potential to elucidate specific features of heart failure-related LV dysfunction.

The tertiary objective is to establish the foundation for a unified, standalone, and clinically deployable framework for comprehensive, automated, and clinician-friendly analysis of LV hemodynamics based on 4D Flow MRI. Internal testing, benchmarking, and structured evaluation by clinical end-users with prior 4D Flow MRI research experience will be conducted to collect feedback and guide further development and clinical translation.

The study includes a retrospective and a prospective arm, addressing methodological development, clinical validation, and translational implementation of advanced MRI-based analysis tools.

Within the retrospective arm, a database of short-axis cine balanced steady-state free precession (bSSFP) images of the LV will be retrieved retrospectively and anonymized prior to analysis. The dataset will be divided into a training set (approximately 75% of cases) and a test set (approximately 25%).

For all MRI datasets included in the training set, LV endocardial contours will be delineated throughout the cardiac cycle, employing semi-automatic segmentation tools (Medviso Segment) with manual corrections applied as necessary to ensure accuracy.

The training dataset, together with the corresponding ground-truth LV endocardial segmentations, will be used to train a deep learning convolutional neural network (CNN), e.g., a ResNet architecture, for the automatic delineation of LV endocardial contours from short-axis cine images.

The remaining test set will be used to evaluate the performance of the trained CNN. Automatically generated LV endocardial contours will be compared with the corresponding manual segmentations to benchmark segmentation accuracy and robustness.

As part of the prospective arm of the study, enrolled patients will undergo non-contrast 4D Flow MRI acquisition prior to the scheduled intervention for transcatheter aortic valve replacement (TAVR).

MRI examinations will be acquired on a 1.5 Tesla scanner (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany). Standard bSSFP cine sequences with retrospective ECG gating will be used to acquire short-axis and long-axis LV views (slice thickness 8 mm, no inter-slice gap, 30 reconstructed cardiac phases). In the same imaging session, 4D Flow MRI data will be acquired using a prototype time-resolved three-dimensional gradient echo sequence with three-directional velocity encoding. A parasagittal-oriented field of view covering the entire LV will be employed. Acquisitions will be performed during free breathing with retrospective ECG triggering and adaptive respiratory gating. Scan parameters will comply with the 2023 update of the 4D Flow Cardiovascular Magnetic Resonance Consensus Statement. No contrast agent will be administered. Non-invasive continuous blood pressure and relevant hemodynamic parameters will be also recorded from the finger arterial pressure waveform using a clinical-grade device (e.g., Finapres® NOVA).

During the TAVR procedure, simultaneous invasive blood pressure measurements will be obtained in the ascending aorta and in the LV using two pigtail catheters, both before and after valve implantation, as required by standard clinical practice during TAVR, thus without impacting on the clinical workflow.

No variations in prophylaxis or follow-up procedures are expected compared with standard clinical practice. Therefore, no additional physical, psychological, or social risks or direct benefits are associated with participation in the study.

Post-processing of 4D Flow MRI data will be performed using a dedicated workflow incorporating in-house software written in Matlab (The MathWorks Inc., Natick, MA, USA), based on prior experience. Processing steps will include correction for eddy currents and velocity aliasing. Intracavitary LV velocity fields will be extracted using both an in-house semi-automatic segmentation tool and the automated deep learning-based LV masking tool developed in the retrospective arm.

Extracted velocity fields will be used to evaluate LV blood flow energetics, including kinetic energy and viscous energy dissipation, flow component subdivision, intraventricular pressure gradients, and intracavitary flow-mediated hemodynamic forces (HDFs). The resulting HDF vectors will be decomposed into basal-apical, septal-lateral, and inferior-anterior components, and their root mean square (RMS) values will be computed over the cardiac cycle.

Non-invasive continuous blood pressure signals derived from finger arterial waveforms will be combined with MRI-derived LV volumes to estimate non-invasive LV pressure-volume (PV) loops using an established methodology. In parallel, invasive LV pressure data acquired during the TAVR procedure via pigtail catheter will be used to generate catheter-based PV loops. Non-invasive MRI-based PV loop parameters will be quantitatively compared with catheter-based measurements using clinically relevant LV indices.

Intracavitary pressure gradients will be computed from the Navier-Stokes equations to derive HDF vectors for each cardiac frame by integrating pressure gradients over the entire LV volume. HDF vectors will be projected onto three orthogonal anatomical directions (basal-apical, septal-lateral, inferior-anterior), and root mean square (RMS) values of each component calculated over the cardiac cycle. Simplified HDFs will be computed using a recently developed method that does not require 4D Flow MRI data and they will be systematically compared with standard 4D Flow MRI-based HDFs on a patient-specific basis. For validation purposes, both simplified and 4D Flow-derived HDFs will also be compared with corresponding measurements obtained using the prototype HDF tool available in the commercial Medis Suite MR software.

Conventional MRI-derived parameters routinely used to assess LV function, including end-diastolic volume, end-systolic volume, and global longitudinal strain, will also be collected.

Demographic variables (e.g., age and sex), relevant clinical characteristics, administered medical therapy, procedural TAVR data, and concomitant medications will be collected and anonymized prior to analysis.

No follow-up visits or assessments are planned as part of this study.

A standalone application for intracardiac 4D Flow MRI analysis will be designed following an incremental build development model. The process will focus on defining the system requirements specification (SyRS), software requirements specification (SRS), and overall software architecture.

A critical evaluation of existing state-of-the-art 4D Flow MRI workflows used in clinical research and of prototype tools available in commercial software (e.g., Medis Suite MR) will be conducted to identify strengths and limitations. Development will proceed through successive stages, with each stage implementing a specific functional component of the overall framework.

The standalone application will build upon existing in-house Matlab-based algorithms but will be implemented in a compiled programming language, allowing execution without requiring pre-installed software on the end-user's system. Development will be supported by a specialized consulting service.

A preliminary version of the standalone application will undergo internal testing, benchmarking, and structured evaluation by a small group of clinical end-users with prior experience in 4D Flow MRI research. Users will independently test the software on predefined use cases. Clinical feedback, criticisms, and suggestions for further development will be collected through structured questionnaires or interviews and used to guide subsequent refinement and clinical translation.

Inclusion Criteria:

• Adult patients (age > 18 years old);
• Diagnosis of severe AS defined according to ESC guidelines with indication to TAVR;
• Severe aortic stenosis both in normal/high flow status and in low flow status;
• Signed informed written consent.

Exclusion Criteria:

• Contraindication to cardiac MRI due to previous implant with ferromagnetic components;
• Poor MRI quality impairing image post-processing;
• Claustrophobia;
• Unwilling to sign the informed consent.