1357197 - Implementing a validated AI tool for automated ventricular function (AI-CMRQC) in a new institution: Impact on LV volume quantification.

Implementing novel AI-based aIgorithms in institutions requires careful consideration. Our quality-controlled, fully-automated method for cardiac function analysis from cine CMR has been extensively trained and validated, including external data from over 12 different institutions^1,2. Evenso, interinstitutional differences in segmentation strategies, training and habits impact volumetric assessment. We investigate how implementation of AI-CMR^QC in new institution impacts ventricular volume quantification.

Methods:

AI-CMR^QC was implemented in a regional district hospital in the Netherlands.

We randomly selected 250 CMR exams, that were acquired and analysed between 1^st January 2018 and 31^st July 2021 from the PACS storage system. We compared the original reported values for LV ventricular assessment with the ones obtained using AI-CMR^QC using Pearson’s correlation and Blant-Altman plots. RV assessment was reported in < 5% of the CMR cases and was therefore not included in this analysis, but a more extensive analysis is currently ongoing. Data or reporting staff of the institution was not previously involved in the development of AI-CMR^QC.

Results:

Pearson’s correlation coefficients were good for all measures of LV volume; LVEDV r=0.96, LVESV r=0.98, LVSV r=0.89, LVEF r=0.90. Mean difference between reported values and AI-CMR^QC was -5.31 mL for LVEDV, -1.02mL for ESV, -4.32mL for SV, and -1.11 % for LVEF (see Blant Altman and Scatter plots in Figures 1).

Conclusion:

There was good to excellent correlation between LV volumes obtained using AI-CMR^QC  compared to those obtained during routine manual clinical reporting, with small mean differences between the two methods. Some individual variability was and could reflect different segmentation strategies. Implementation of AI-CMR^QC did not meaningfully affect LV volume quantification.