inHEART Models

inHEART MODELS is a medical imaging software suite designed to assist qualified medical professionals in cardiac treatment planning by processing CT and MR images to generate interactive 3D anatomical models. The system employs a combination of AI-driven automated segmentation for adult CT images (utilizing a UNet machine learning algorithm trained on 796 cases), along with manual and semi-automated tools for model generation, visualization, analysis, and annotation to support clinical decision-making.

inHEART MODELS is a comprehensive suite of medical imaging software tools designed to assist qualified medical professionals, including physicians, cardiologists, radiologists, and technicians, in the reading, interpretation, reporting, and treatment planning for cardiac conditions. The system processes DICOM-compliant medical images from CT and MR scanners to generate detailed 3D models of patient anatomy. It comprises three main software components: inHEART MODELS AI, which uses a machine-learning (UNet) based approach for automatic 3D modeling and pre-processing of adult CT images; inHEART MODELS Shaper, a standalone software for generating digital 3D heart models from CT and/or MR images through manual and semi-automated processing; and inHEART MODELS Explorer, a web-based platform for displaying, reviewing, analyzing, annotating, and exporting these 3D cardiac models. The AI module was trained on 796 anonymized CT cases from US and European institutions, employing data augmentation, regularization methods, specific loss functions, and optimization methods over 1000 epochs. The performance of the inHEART MODELS AI module, particularly its ability to provide initial segmentations for review and refinement, was validated using a separate dataset of 100 CT cases, ensuring independence from the training data. These AI-assisted segmentations, after expert review and editing using the inHEART MODELS Shaper, were compared against a ground truth established from expert-approved segmentations. The validation focused on metrics like Dice coefficient, Average Symmetric Surface Distance (ASSD), and volumetric differences for various cardiac structures including the aorta, left atrium, left ventricle endocardium and epicardium, pulmonary artery trunk, right atrium & inferior vena cava, and right ventricle endocardium and epicardium. Results for these revised AI-based segmentations demonstrated high accuracy, with an overall average Dice score of 0.94, an average ASSD of 1.17mm, and average volume variations of 9.18mL (7%), meeting predefined acceptance criteria and indicating the system's capability to produce accurate anatomical models to support therapeutic interventions.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the performance of the inHEART MODELS AI module vary with different CT image qualities and across diverse patient populations not extensively represented in the initial 796-case training dataset?; What is the typical time saving and impact on inter-observer variability for cardiac structure segmentation when using the inHEART MODELS AI module followed by expert review, compared to fully manual or semi-automated methods?; Are there specific complex cardiac anatomies or post-intervention scenarios where the inHEART MODELS AI requires significantly more manual correction by expert users, and what is the learning curve for users to efficiently refine these AI-generated segmentations?; How does the web-based inHEART MODELS Explorer facilitate collaborative review and annotation of 3D cardiac models among multidisciplinary teams, and what are its data security features for handling sensitive patient information?; What are the plans to expand the training dataset for the inHEART MODELS AI to improve its robustness, and are there ongoing efforts to validate its utility for MR images or for pediatric patients, given its current indication for adult CT images only?