Assessment of commercially available artificial intelligence software for differentiating hemorrhage from contrast on head CT following thrombolysis for ischemic stroke

 

And you don't know about much faster ways to determine infarct or bleed? 

My God THE FUCKING INCOMPETENCE IN STROKE IS MIND-BOGGLING!

Hats off to Helmet of Hope - stroke diagnosis in 30 seconds; February 2017 

Smart Brain-Wave Cap Recognises Stroke Before the Patient Reaches the Hospital

 October 2023

And then this to rule out a bleeder.

New Device Quickly Assesses Brain Bleeding in Head Injuries - 5-10 minutes April 2017

The latest here:

Assessment of commercially available artificial intelligence software for differentiating hemorrhage from contrast on head CT following thrombolysis for ischemic stroke

Ayden L. Olsen¹Daniel Thomas Ginat^2*

• ¹Pritzker School of Medicine, The University of Chicago, Chicago, IL, United States
• ²Department of Radiology, Section of Neuroradiology, The University of Chicago, Chicago, IL, United States

Background: In patients who have undergone ischemic stroke therapy, retained iodine-based contrast can resemble acute intracranial hemorrhage (ICH) on standard computed tomography (CT). The purpose of this study is to determine the accuracy of commercially available artificial intelligence software for differentiating hemorrhage from contrast in such cases.

Methods: A total of 45 CT scans analyzed by Aidoc software that also included dual-energy iodine subtraction maps from dual energy CT from 23 unique patients (12 male, 11 female, age range 30–99 years, mean age 67.6 years, standard deviation 18.5 years) following recent ischemic stroke therapy were retrospectively reviewed for the presence of hemorrhage versus retained contrast material.

Results: The sensitivity and specificity of the model in detecting acute intracranial hemorrhage as opposed to contrast were 51.7 and 50.0%, respectively. The positive and negative predictive values were 65.2 and 36.4%, respectively.

Conclusion: The current Aidoc software is not optimized for differentiating between acute hemorrhage and retained contrast on CT. This justifies the development of a more robust artificial intelligence model trained to differentiate between ICH and iodine contrast based on both DECT and standard CT images.

Introduction

Acute intracranial hemorrhage (ICH) after intravenous thrombolytic therapy for ischemic stroke is rare but potentially life-threatening depending on the size and location of the bleed. Rapid identification and management are essential to achieving positive patient outcomes (1). Following ischemic stroke, patients may receive acute interventions such as intravenous tissue plasminogen activator (tPA) administration (e.g., alteplase or tenecteplase) or mechanical thrombectomy to restore cerebral blood flow. Diagnostic imaging prior to treatment typically involves the use of an iodine-based contrast agent to evaluate vascular occlusion, collateral flow, and perfusion status. However, blood–brain barrier (BBB) leakage can lead to retention of contrast material, which may persist in subsequent non-contrast imaging, complicating the differentiation between retained contrast and acute ICH. Post-treatment imaging, typically performed without the administration of an iodine contrast agent, aims to detecting complications such as ICH, but retained iodine contrast from pre-treatment imaging may appear as hyperdense foci on standard head CT, mimicking hemorrhage. Previous studies estimate that iodine extravasation may account for up to 84% of hyperdense foci seen on follow-up scans after non-mechanical thrombolysis for ischemic stroke (2, 3). Thus, retained contrast can present a challenge in post-treatment imaging, as it is difficult to differentiate from acute ICH.

Dual-energy computed tomography (DECT) provides a solution to this diagnostic challenge by utilizing two unique photon energy spectra to better distinguish materials with varying attenuation properties (4). With post-processing, DECT exams can be used to generate an iodine overlay map (IOM), enabling the creation of virtual non-contrast (VNC) images (5–8). These VNC images subtract densities corresponding to iodine contrast from the enhanced images, significantly aiding in the differentiation between retained iodine contrast and acute ICH. Because of its ability to generate IOMs and VNC images, DECT achieves nearly 100% accuracy in distinguishing between retained iodine contrast and hemorrhage in scenarios of post-ischemic stroke therapy (9). While VNC images generated by DECT may slightly differ in the exact attenuation values as compared to true non-contrast CT (NCCT), multiple studies have shown that these differences are relatively minor and that VNC images can be used reliably in diagnostic scenarios (10–12).

With the increasing use of artificial intelligence (AI) as a useful tool in the field of radiology over the past several years, many triage AI models have been developed and implemented in various hospitals across the world. While some of these AI models perform at very high accuracy, others are underdeveloped or not trained to a level at which they can be consistently reliable in clinical situations. In the setting of post-stroke therapy CT imaging, an AI model able to differentiate between retained contrast and acute hemorrhage would enable the flagging of cases with suspected ICH, allowing for more rapid identification and intervention, leading to better patient outcomes.

Recently, Aidoc (Tel Aviv, Israel) has created an artificial intelligence (AI) model to detect acute ICH on standard CT images, flagging images with suspected hemorrhage for further review by a trained radiologist. Despite the promisingly high performance reported (13), this model does not appear to be optimized for detecting the difference between ICH and retained contrast material on conventional CT exams. We have previously shown that the implementation of an automatic flagging system decreases scan view delay time and expedited diagnosis of urgent conditions (14). Because of the critical need to quickly identify and assess potential cases of ICH, it is important to assess the ability of the Aidoc model to accurately distinguish between ICH and retained iodine contrast, and, if it cannot, to develop a novel model that can do so with more generalizable accuracy.

In this study, we use the Aidoc model to predict contrast versus hemorrhage for 45 CT exams showing hyperdensities confirmed to be either retained contrast or acute hemorrhage. Through this analysis, we show that performance of current AI models in distinguishing between retained contrast and acute ICH on post-stroke therapy CT scans is insufficient, and a more robust model needs to be trained and validated.

More at link.