Objective:

Evaluate the performance of an automated detection and quantification tool for identification and severity assessment of amyloid-related imaging abnormalities (ARIA).

Background:

Amyloid beta-directed monoclonal antibody therapies may be associated with ARIA-E(edema/sulcal effusion), and ARIA-H (hemorrhage and superficial siderosis). Studies have shown that clinicians find it challenging to identify or evaluate ARIA. As treatment decisions are informed by ARIA presence and severity, an assistive automated detection and diagnosis tool may help to improve radiological reading.

Design/Methods:

ARIA-E and ARIA-H were identified on 2D T2 FLAIR and T2*-GRE images, respectively. icobrain aria, an automated detection and diagnosis tool, was trained on data from the EMERGE clinical trial (475 FLAIR, 326 T2*GRE images) using U-net-style convolutional neural networks. Ground truth was established by three expert neuroradiologists. 16 radiologists evaluated ARIA for 199 cases, assisted and unassisted by icobrain aria. Sensitivity and specificity of assisted and unassisted readings relative to ground truth were compared for three categories of ARIA (none, mild, moderate-or-severe).

Results:

Sensitivity of detection for ARIA-E and ARIA-H increased significantly when reading was assisted by icobrain aria compared to unassisted reads (from 47.2% to 70.2% for mild ARIA-E; from 83.7% to 95.3% for moderate-or-severe ARIA-E; from 60.0% to 71.0% for mild ARIA-H; from 85.5% to 94.4% for moderate-or-severe ARIA-H). In ARIA-free cases, assisted reads had specificity of 83.0% for mild ARIA-E, 80.3% for mild ARIA-H, 94.0% for moderate-or-severe ARIA-E, and 99.4% for moderate-or-severe ARIA-H.

Conclusions:

By increasing sensitivity of ARIA detection by radiologists, icobrain aria may help improve safety monitoring for patients on amyloid beta-directed monoclonal antibody therapies. Additionally, icobrain aria may help to bridge the gap between expert and non-expert readers.