HepaFatSmart (V2.0.0)

HepaFatSmart is a software device intended for non-invasive, quantitative assessment of liver fat (VLFF, PDFF) and steatosis grade using AI-driven analysis of MRI data. Its methodology relies on a convolutional neural network for automated liver ROI detection and subsequent algorithmic calculation of fat fractions based on the differential resonance frequencies of hydrogen nuclei in water versus triglyceride fat, aiming to support clinical diagnosis and management of liver conditions.

HepaFatSmart is a Software as a Medical Device (SaMD) designed for the quantitative measurement of liver fat, specifically volumetric liver fat fraction (VLFF), proton density fat fraction (PDFF), and steatosis grading. It operates by automatically analyzing Magnetic Resonance Imaging (MRI) datasets, which users upload to a secure portal; its core methodology utilizes AI, specifically a convolutional neural network (CNN), to predict a suitable liver region of interest (ROI) and remove undesired components like artifacts and major blood vessels. From this ROI, it calculates an 'Alpha' value, a magnetic resonance parameter reflecting the ratio of triglyceride fat proton density signal to total liver proton density signal, which is then converted to VLFF, PDFF, and steatosis grade. The system incorporates an automated image quality control (IQC) module for MRI protocol adherence and an algorithmic assessment for excessive liver iron, ensuring reproducibility with a locked-down AI model post-training. Performance validation studies demonstrated that HepaFatSmart achieves high repeatability and produces results substantially equivalent to the manual HepaFat-Scan method, while being superior to a previous AI version, HepaFat-AI. A repeatability study (n=41) showed HepaFatSmart had slightly better bias and repeatability coefficients than HepaFat-Scan and achieved 100% reproducibility (zero VLFF difference) on re-analysis of the same datasets. A larger validation study (n=281 after IQC filtering) comparing HepaFatSmart to HepaFat-Scan confirmed strong agreement, revealing a statistically significant but clinically unlikely significant bias of 0.2% VLFF. It demonstrated high sensitivity and specificity for clinically relevant VLFF thresholds (e.g., for VLFF >4.1%, sensitivity was 100.0% and specificity 98.6%). The output, intended for interpretation by a trained physician, comprises automatically generated PDF and DICOM reports detailing VLFF, PDFF, steatosis grade, associated confidence intervals, normal ranges, images of the analyzed MRI slice, the AI-predicted ROI, and a liver fat distribution map.

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 AI-driven ROI selection in HepaFatSmart compare in accuracy and robustness to experienced radiologist manual ROI selection across diverse liver morphologies, patient populations (e.g., pediatric, obese), and varying MRI image qualities not extensively covered in the validation set?; What is the specific algorithm and performance (sensitivity/specificity) of HepaFatSmart's excessive liver iron assessment module across different degrees of iron overload, and how does it quantitatively adjust or flag results in such cases?; Given the statistically significant, though small (0.2%), bias in VLFF compared to HepaFat-Scan, what is the potential cumulative impact of this bias in longitudinal monitoring of patients with fatty liver disease, particularly those near clinical decision thresholds?; How does the HepaFatSmart system handle MRI datasets with significant artifacts (e.g., motion, chemical shift, susceptibility) that might pass the IQC module, and what is the reported failure analysis rate or need for manual review in real-world clinical practice?; What is the comparative diagnostic performance and cost-effectiveness of HepaFatSmart against other non-invasive liver fat quantification methods, such as transient elastography (FibroScan with CAP) or other MRI-PDFF software solutions, in routine clinical use for diagnosing and managing NAFLD?