Automated abstraction of myocardial perfusion imaging reports using natural language processing

“Global” cardiac atherosclerotic burden assessed by artificial intelligence-based versus manual segmentation in 18F-sodium fluoride PET/CT scans: Head-to-head comparison

“…The current CNNs were trained beforehand to segment the heart automatically. The manual segmentation was performed by two image analyzers (RP and SS) in NaF-PET/CT scans, which were obtained as part of the prospective 'Cardiovascular Molecular Calcification Assessed by 18 F-NaF PET/CT' (CAMONA) study conducted at Odense University Hospital, Odense, Denmark, in 2012-2014. 9 The heart was segmented in a way to encompass both the left and right atria and the ventricles.…”

“…In nuclear cardiology, such methods have been used to facilitate attenuation correction 17 and automate reports of myocardial perfusion imaging. [18][19][20][21] However, very few have used artificial intelligence to optimize the interpretation of PET imaging. 22 We have done this to facilitate assessment of early atherosclerosis in the large arteries and changes therein over a period of 2 years.…”

“…1 We set out to examine if we could train and test an artificial intelligence (AI)-based, socalled convolutional neural network (CNN), to provide a reliable score expressing global cardiac atherosclerotic burden in less than a minute. We did so in 18 F-sodium fluoride (NaF)-PET/CT scans assuming that a score with this tracer might be of significant clinical interest. It would represent early atherosclerosis in the entire coronary arterial tree, including transmural and minor arteries subtending the most ischemia-prone parts of the myocardium, whereas traditional assessment with more common noninvasive techniques like the Agatston score by low-dose CT, CT angiography, optical coherence tomography or magnetic resonance imaging depict late-stage atherosclerosis in epicardial arteries.…”

Automated abstraction of myocardial perfusion imaging reports using natural language processing

“…This is further compounded by natural factors such as fatigue, distraction, and stress which interfere with the consistency and accuracy of human abstraction of data. Clinical natural language processing (cNLP) as described in this study by Zheng et al 2 , in this issue of the journal, allows for abstraction and organizing of the data from free text. cNLP is trained to understand the unique medical language, abbreviations, and common medical jargon and output this text into structured information which renders itself for statistical analysis and perhaps can then be used to guide clinical decisions.…”

“…Myocardial perfusion imaging (MPI) with SPECT or PET is a well-established non-invasive diagnostic and prognostic tools for the evaluation of patients with known or suspected coronary artery disease (CAD) 6,7 . The study by Zheng et al 2 is not directed toward interpretation of MPI images. This program evaluated the efficacy of cNLP in abstracting data and key variables, which can be fed directly to a data base for statistical analysis from unstructured and unanalyzable free hand reports of MPI data.…”