A novel “Transformer”-based deep learning (DL) model improves diagnostic accuracy of plaque erosion by optical coherence tomography (OCT) in comparison to a convolutional neural network (CNN) DL model, according to a new study. Sangjoon Park, MD, of the Korea Advanced Institute of Science and Technology, Daejon, South Korea, Makoto Araki, MD, PhD, of Massachusetts General Hospital and Harvard Medical School, Boston, and colleagues reported these findings in a manuscript published online Monday and in the Oct. 24 issue of JACC: Cardiovascular Interventions. Acute coronary syndrome (ACS) may be due to one of three causes: plaque rupture, eruptive calcified nodule or plaque erosion. Plaque erosion can only be detected using intracoronary imaging and with that, remains difficult to identify by the average reader. Its detection may have important clinical implications, as recent studies have suggested plaque erosion may be treated with a course of antithrombotics rather than stent implantation. DL, a branch artificial intelligence, is utilized increasingly in the medical field to improve diagnostic accuracy. A “Transformer”-based DL model uses multiple frames to make a diagnosis rather a single frame used in current models. Park, Araki and colleagues, therefore, set out to investigate the ability of Transformer DL model to improve diagnostic accuracy of plaque erosion. They analyzed data from patients presenting with ACS who had pre-percutaneous coronary intervention OCT imaging of the culprit lesion in the PREDICTOR (Identification of Predictors for Coronary Plaque Erosion in Patients with Acute Coronary Syndrome) study dataset, and they externally validated using the EROSION study dataset. A total of 237,021 cross-sectional OCT images from 581 patients were used for training and internal validation, and 65,394 images from 292 patients from the other dataset were used for external validation. The Transformer model showed superior performance to that of the CNN model in the frame level (area under the curve [AUC] of 0.94 compared with 0.85) and the lesion-level (AUC 0.91 compared with 0.84 in the external validation). Sensitivity to identification of plaque erosion was 91.9%, the specificity was 81.6%, and the positive predictive value was acceptable at 67.5%. When comparing findings to interpretation by readers, the model outperformed inexperienced readers but was inferior to expert readers. Notably, novice readers aided by the model brought their performance closer to that of an experienced reader. Robert Avram, MD, MSc, and Guillaume Marquis-Gravel, MD, MSc, of the Montreal Heart Institute at the University of Montreal, wrote an accompanying editorial comment. They commend the authors for applying this new technology in machine learning to a clinically useful task. They did also highlight several limitations. First, OCT was performed on a operator discretion basis biasing the validation data, therefore possibly limiting the applicability in the real world setting. Another limitation is that the Transformer model’s positive predictive value was lower than that of the expert reader—20.0% to 32.5% would be false positive plaque erosions – but sensitivity was quite high. Therefore, the DL model, at this time, serves best as a rule-out test, but confirmatory diagnosis by an expert reader is needed when plaque erosion is detected. “Overall, this study is an important step in understanding how [artificial intelligence] can automatically recognize and improve plaque erosion diagnosis on OCT,” the editorialists concluded. “Such preliminary work can pave the way for an easy-to-use and automatic method for OCT interpretation in ACS and personalized management of plaque lesions.” Sources: Park S, Araki M, Nakajima A, et al. Enhanced Diagnosis of Plaque Erosion by Deep Learning in Patients With Acute Coronary Syndromes. JACC Cardiovasc Interv 2022;15:2020–2031. Avram R, Marquis-Gravel G. Automated Artificial Intelligence–Guided Diagnosis of Plaque Erosion: Ready for Prime Time? JACC Cardiovasc Interv 2022;15:2032–2034. Image Credit: wladimir1804 – stock.adobe.com