Pulmonary Embolism

Purpose	Detection and assessment of pulmonary embolism on CT imaging
Tag(s)
Panel	Thoracic
Define-AI ID	19080011
Originator	Scott J. Adams
Lead	Scott J. Adams
Panel Chair	Eric J. Stern
Panel Reviewers	Thoracic Panel
License	Creative Commons 4.0
Status	Published
RadElement Set	RDES80

                               

Value Proposition

Pulmonary embolism (PE) is the third most common acute cardiovascular disease after myocardial infarction and stroke. Clinical presentation of patients with PE typically includes dyspnea, chest pain (particularly pleuritic or sometimes dull), or cough; however, clinical presentation can range from being asymptomatic to sudden death, and urgent diagnosis is critical. PE may also be an incidental finding on CT studies of the chest. At some centers there may be an interval of multiple days between when studies such as outpatient cardiac studies are performed and when they are interpreted, making it important that critical findings such as PE are brought to the attention of a radiologist in a timely manner.
CT pulmonary angiography has become the most common imaging modality to assess for PE. The hallmark feature of acute PE on CT imaging is an intraluminal filling defect which has a sharp interface with intravascular contrast material. Assessment of features associated with PE, including right ventricular heart dysfunction, is critical to assess the risk of circulatory collapse, and measures such as right ventricular to left ventricular (RV/LV) ratio have been demonstrated to be associated with short-term mortality. Quantitative or semi-quantitative estimates of clot burden may be helpful to assess the evolution of PE and monitor effects of treatment. Features of chronic PE include complete occlusion of a pulmonary artery which is often smaller than adjacent patent vessels, a thickened and smaller caliber pulmonary artery with some contrast-opacification, and intraluminal webs or bands. In addition, CT findings of both acute and chronic PE can occur together and are not mutually exclusive.

Artificial intelligence for detection of PE would allow for prioritization of positive studies for interpretation by a radiologist and/or a radiologist being notified of critical findings. In addition to aiding in clot detection, automatic calculation of values relevant to interpretation of CT pulmonary angiography studies, such as the RV/LV ratio and overall clot burden, may increase radiologists’ efficiency in reporting studies and allow for valuable information to be easily communicated to referring clinicians.

Narrative

A 56-year-old female with breast cancer presents for a CT pulmonary angiography study following assessment at an outpatient medical oncology appointment at which she complained of increasing shortness of breath. CT pulmonary angiography images are analyzed by the AI engine, which determines there is a filling defect most proximally in the right upper lobar pulmonary artery with associated right heart strain. An alert message is sent to PACS prioritizing the study for interpretation by a radiologist, and information indicating the location of the PE, most proximal extent, total clot volume, and RV/LV ratio is sent from the AI engine to PACS.
A 59-year-old man with bladder cancer presents for outpatient CT chest, abdomen, and pelvis for restaging of non-Hodgkin lymphoma. The CT images are analyzed by the AI algorithm, which determines that multiple subsegmental filling defects are present bilaterally. There is no CT evidence of right ventricular strain. The study is prioritized on the CT reading worklist with a message to alert radiologists of the presence of incidental pulmonary emboli.

Workflow Description

Images are sent from the modality workstation to PACS and the AI engine. The images are analyzed by the AI engine. The AI engine detects PE and assesses the level of proximal extension, clot burden, and right heart dysfunction. An alert message is sent to PACS from the engine to prioritize the study for review and/or notifies a radiologist about possible critical findings requiring review. Information indicating the location of the PE, most proximal level of extent, total clot volume, and RV/LV ratio is sent from the AI engine to PACS. Additionally, image(s) indicating the predicted location(s) of PE are sent to PACS.

Procedure(s)	{CT pulmonary angiography, CT chest w/ contrast, CT cardiac w/contrast, CT cardiac w/o contrast, CT thoracic aorta w/ contrast, CT thoracic aorta w/o contrast}
Slice thickness	Variable
Age	(0, +∞)
Adequacy of study	{non-diagnostic study, diagnostic to the level of the main pulmonary arteries, diagnostic to the level of the lobar pulmonary arteries, diagnostic to the level of the segmental pulmonary arteries, diagnostic to the level of the subsegmental pulmonary arteries}
Presence of pulmonary embolism	{present, absent, indeterminate}
Level of proximal extension	{main pulmonary artery, interlobar pulmonary artery, lobar pulmonary artery, segmental pulmonary artery, subsegmental pulmonary artery}
Laterality	{bilateral, right, left}
Lobar involvement	{right upper lobe, right middle lobe, right lower lobe, left upper lobe, left lower lobe}
Image reference	{manual or semi-automated segmentation of filling defects, or image and series numbers demonstrating filling defects}
Clot Burden	{Qanadli score, Mastora score, fully quantitative measures of total clot volume}
RV/LV ratio	[0, +∞) Clinical Note: Many sources consider normal RV/LV ratio to be between zero and one.
Temporality	{acute, acute on chronic, chronic, indeterminate} Note: Datasets to test algorithm performance should include common mimics of PE, including respiratory motion, image noise, flow-related artifacts, "stair step" artifacts, and mucus plugs, which are potential sources of false positives.

Inputs

DICOM Study

Primary Outputs

Adequacy of the study

Secondary Outputs

Level of Proximal Extension

Laterality

Lobar Involvement

Image Reference/Saliency Map

Clot Burden

RV/LV Ratio

Temporality

A variety of technical factors limit accurate assessment for PE on CT pulmonary angiography studies, and often studies are reported as being indeterminate beyond a certain level of pulmonary vasculature (e.g. segmental or subsegmental arteries). Artificial intelligence for image denoising/artifact reduction may assist radiologists in interpreting CT pulmonary angiography studies, as well as aid in subsequent computer-aided assessment of PE.

Multiple features may be used to assess right heart dysfunction, including ratio of right ventricle diameter to left ventricle diameter, morphology of the interventricular septum, contrast medium reflux into the inferior vena cava, ratio of main pulmonary artery diameter to ascending aorta diameter, superior vena cava diameter, and azygos vein diameter. Artificial intelligence tools may be used to individually assess each of these features, or make a global prediction about right heart dysfunction.

Automatic tracking of clot volume and extent over time may aid in the assessment of chronic PE. Additionally, automatic insertion of a description of findings into reporting software may increase radiologists’ efficiency.

No known related public datasets at this time, please alert us if you know of any.