FerriSmart Analysis System

FerriSmart is an AI-driven software application that non-invasively quantifies liver iron concentration by analyzing R2 signal decay from specific MRI sequences, offering a diagnostic tool for managing iron overload conditions. Its methodology leverages convolutional neural networks for automated image analysis and an established calibration curve to convert MRI-derived R2 signals into clinically actionable liver iron measurements, with performance validated against clinically relevant thresholds and a predicate device.

FerriSmart is a stand-alone software application designed to quantitatively measure liver iron concentration (LIC) from Magnetic Resonance Imaging (MRI) scans. Its primary purpose is to aid in diagnosing and monitoring individuals with confirmed or suspected systemic iron overload, particularly those with transfusion-dependent thalassemia or sickle cell disease receiving blood transfusions, and to assist in the identification and monitoring of non-transfusion-dependent thalassemia patients undergoing therapy with deferasirox. The system utilizes a specific multi-slice, spin-echo MRI protocol to acquire abdominal image data. This data is then automatically analyzed by FerriSmart's Artificial Intelligence (AI)-powered software, which employs convolutional neural networks to determine the signal decay rate (R2) in the liver. An integrated Input Quality Control (IQC) module, an automated algorithm, verifies the correctness of the MRI data acquisition parameters before analysis. The measured R2 value, which characterizes iron loading in the liver, is subsequently transformed by a defined calibration curve to provide a quantitative measure of LIC in vivo, reported in micromoles and milligrams per gram of dry liver tissue. Clinical performance data, derived from studies including 971 datasets, demonstrated good sensitivity and specificity (mostly above 90%) for FerriSmart in predicting LIC results (compared to the predicate FerriScan) across various clinically relevant thresholds, such as 92% sensitivity and 97% specificity at a 7.0 mg Fe/g threshold. Repeatability studies indicated FerriSmart's precision is consistent with FerriScan below 3 mg Fe/g dry tissue and has an acceptable standard error of approximately 9% above this level, which is notably better than liver biopsy. Although a bias exists when compared to FerriScan, especially at LIC values above 3 mg Fe/g (rendering the two systems not interchangeable), the overall diagnostic performance is considered clinically acceptable. The final report includes the analyzed images for review by a radiologist to assist in clinical diagnosis and management decisions.

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 FerriSmart AI's automated image analysis and IQC module compare to the human-interactive analysis of the predicate device (FerriScan) in terms of processing time and handling of suboptimal image quality?; Given that FerriSmart and FerriScan are not considered interchangeable despite using the same calibration curve, what are the specific magnitudes and clinical implications of the bias observed at different LIC levels, and how should clinicians account for this when transitioning patients previously monitored with FerriScan?; What specific training data was used for the convolutional neural networks in FerriSmart, particularly concerning the diversity of MRI scanner vendors, field strengths, and patient populations, and how does this impact its generalizability across different clinical settings?; The repeatability study mentions a standard error of approximately 9% for FerriSmart above 3 mg Fe/g. How does this precision compare to the typical longitudinal changes in LIC observed in patients undergoing chelation therapy, and is it sufficient to reliably detect clinically significant responses?; What are the failure modes of the automated Input Quality Control (IQC) module, and what is the recommended workflow if the IQC module rejects images or if a radiologist suspects an erroneous automated analysis despite IQC approval?