P1505: using artificial intelligence neural networks to obtain automated liver iron concentration measurements using magnetic resonance imaging – a multi-scanner validation study

Background: Because of the limitations of serum ferritin and liver biopsy for measuring liver iron concentration (LIC), magnetic resonance imaging (MRI) methods have been developed. Reliable use of these methods requires extra training of radiologists or outsourcing of the data analysis. While the cost of outsourcing data analysis is acceptable to many patients or providers in high income countries, the majority of hemoglobinopathy patients reside in low-income countries. Therefore, automated LIC measurements using artificial intelligence (AI) or deep-learning-assessed (DLA) algorithms could provide the solution for globally affordable and reliable patient monitoring of LIC. Aims: The aim of this study was to evaluate the performance of an automated deep-learning-based medical device (DLA R2-MRI) for measuring liver iron concentration (LIC) from MRI using an independent multi-scanner dataset. Methods: The tested device was assessed prospectively on 1395 eligible consecutive MRI datasets from 66 different scanners submitted for expert manual analysis using spin-density projection assisted (SDPA) R2-MRI (the reference standard) between August 2017 and July 2020. The aetiologies for iron overload reported by the radiologists submitting the image data were thalassaemias (498), hereditary haemochromatosis (161), sickle cell disease (145), MDS (11), other (309), unknown (271). The bias and limits of agreement between the automated and manual measurements of LIC were assessed. In addition, diagnostic performance was assessed using sensitivity and specificity. Results: The distribution of LIC values measured by the reference method was 0-1.8 mg Fe/g (17.6%), 1.8-3.2 mg Fe/g (16.7%), 3.2 – 5.0 mg Fe/g (14.8%), 5.0 – 7.0 mg Fe/g (11.4%), 7.0-15.0 mg Fe/g (18.6%), >15.0 mg Fe/g (21.0%). Automated LIC results from the DLA R2-MRI had geometric mean ratios to manual results from SDPA R2-MRI of 0.98 (95% CI 0.94 – 1.01) below 3 mg Fe/g dry tissue and 0.93 (95% CI 0.92 – 0.95) above 3mg Fe/g dry tissue were recorded. The sensitivities and specificities of the automated system for predicting LIC values above clinically relevant thresholds of 3.0, 3.2, 5.0, 7.0, and 15.0 mg Fe/g dry tissue were > 90%. Summary/Conclusion: While there is an overall bias between DLA R2-MRI and SDPA R2-MRI, the bias does not result in unacceptable sensitivities and specificities of DLA R2-MRI for predicting SDPA R2-MRI results above the clinically relevant LIC thresholds. However, the bias between the automated and manual methods indicates that the two techniques should not be used interchangeably.