Search Results

inHEART MODELS comprises a suite of medical imaging software modules that are intended to provide qualified medical professionals with tools to aid them in reading, interpreting, and treatment planning, inHEART MODELS accepts DICOM compliant medical images acquired from a variety of imaging devices, including CT and MR*. The software is designed to be used by qualified medical professionals (including physicians, cardiologists, radiologists, and technicians) and the users are solely responsible for making all final patient management decisions.

• inHEART Models AI software module is indicated for adults only and is designed for CT images only.

inHEART MODELS is a suite of medical image processing software tools that enables 3D visualization and analysis of anatomical structures.

This software suite is composed of three software as a medical device components:

• . inHEART MODELS AI: a medical image processing software for automatic 3D modelling, used to pre-process medical images (acquired only by CT devices).
  This software module uses a machine-learning based approach with the following characteristics:

• Training dataset: 796 cases (3D CT original images and the segmentation । masks) from previously manually performed segmentations (time period 2018-2022); origins of the data are public and private clinical and hospital institutions located in US (40%) and Europe (60%).

• -Training process: Machine learning algorithm (UNet) is trained applying data augmentation (including patch mirroring) and regularization (including Instance Normalization, Dropout, Data Augmentation, Weight Initialization, Leaky ReLU Activation) methods. Loss functions (soft dice loss and binary cross entropy) and optimization methods (stochastic gradient descent) are used during learning over a fixed number of 1000 epochs, each consisting of 250 iterations with a batch size of 2.

• Data anonymized prior to its processing by inHEART team, therefore no । personal data (gender, age, ethnicity) is exploitable. CT scanner manufacturers include Siemens (40%). GE Medical Systems (30%). Toshiba (10%) and other manufacturers (20%) among which Philips and Canon.

• inHEART MODELS Shaper: a standalone image processing software used to . generate digital 3D models of the patient heart from medical images acquired by CT and/or MR devices and;

• inHEART MODELS Explorer: a web-based 3D visualization software that permits . display, review, analysis, annotation and export of the cardiac 3D models generated from inHEART MODELS Shaper.

Specifically, these software modules read DICOM compatible pre-operative CT and MR images acquired by commercially available imaging devices. These images are then processed to generate 3D models of the anatomy to allow qualified medical professionals to display, review, analyze, annotate, interpret, export, and plan therapeutic interventions.

inHEART MODELS software suite also includes two non-device Medical Device Data Systems (MDDS) modules that are only intended to transfer, store and convert formats.

Here's a breakdown of the acceptance criteria and the study proving the device's performance, based on the provided text:

Acceptance Criteria and Study Details for inHEART MODELS

The inHEART MODELS AI software module's performance was evaluated to ensure its automated segmentation capabilities, when reviewed and edited by expert users, yield results comparable to those obtained using the previously cleared predicate device alone.

1. Table of Acceptance Criteria and Reported Device Performance

The acceptance criteria are defined based on the objective of providing an initial segmentation that accelerates the process for expert segmenters, rather than replacing manual expert segmentation entirely. The reported performance metrics are computed between the revised segmentations made by experts based on AI results and the Ground Truth (GT).

Note on Volumetric Analysis: The table in the document provides values for individual structures. The "Overall Reported Device Performance" rows above are calculated as the average/median of the means and medians respectively across all listed structures from the provided table.

Individual structure performance:

2. Sample Size and Data Provenance

• Test Set Sample Size: 100 CT cases.
• Data Provenance (Test Set): Randomly selected from the time period year -2023 to ensure independence from training datasets. The sources are similar to the training dataset, including new clients. The overall geographic origin for training data (which implies similar origins for test data) is US (40%) and Europe (60%).
• Retrospective/Prospective: Not explicitly stated, but the selection from a time period "year -2023" for testing suggests a retrospective approach.

3. Number and Qualifications of Experts for Ground Truth

• Number of Experts: Two external experts.
• Qualifications of Experts: Radiologists. No specific years of experience are detailed, but they are referred to as "external experts" and were tasked with evaluating concordance of manual segmentations.

4. Adjudication Method for the Test Set

The ground truth was established by "Approval by external experts (radiologists) of segmentations previously obtained with the predicate device, that will constitute the Ground Truth (GT)". This implies a consensus or agreement method, but the specific adjudication method (e.g., 2+1, 3+1) is not explicitly stated. It seems the two experts had to agree for the segmentations to form the GT.

5. Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study

An MRMC study was not explicitly described in terms of human readers improving with AI vs. without AI assistance. The study compares the performance of the AI-assisted, expert-reviewed results against the originally expert-segmented predicate device results (Ground Truth). The stated "effect size" is the performance metrics (Dice, ASSD, Volume difference) demonstrating that the AI-assisted workflow, after expert review and editing, is substantially equivalent to the manual predicate workflow. The purpose of the AI module is to accelerate the process for the expert segmenter by providing an initialization.

6. Standalone (Algorithm Only) Performance

Yes, a form of standalone performance was assessed as part of the study: "Comparison of unprocessed outputs of the new automated software module results with the GT to establish the internal validation of the tool performance". However, the specific metrics for this direct AI output (without expert correction) are not separately provided in the summary table. The provided table (under "Performance testing results") shows the metrics after expert review and revision based on AI results.

7. Type of Ground Truth Used

The ground truth (GT) was established through expert consensus/approval of segmentations previously obtained using the predicate device. This is referred to as "Approval by external experts (radiologists) of segmentations previously obtained with the predicate device, that will constitute the Ground Truth (GT)".

8. Sample Size for the Training Set

• Training Dataset: 796 cases (3D CT original images and segmentation masks).

9. How the Ground Truth for the Training Set Was Established

The ground truth for the training set consisted of "previously manually performed segmentations". It is implied these manual segmentations were performed by qualified personnel, similar to how the predicate device was used to establish the test set GT. The document does not specify the number of experts or the exact process (e.g., consensus) for generating the training set ground truth, but it strongly suggests expert-generated manual segmentations.

Ask a specific question about this device