Document resource
Sources of Funding Supported by British Heart Foundation Grants PG/18/14/33562, RG13/14/30314, RE/24/130011, TA/F/20/210001 (London), Academy of Medical Sciences Starter Grants for Clinical Lecturers (REF: SGL030\1012), Innovate UK Advancing Precision Medicine 10069871, National Institutes of Health, R01 HL150608, EPSRC Cambridge Maths in Healthcare (Nr. EP/N014588/1, and Cambridge NIHR Biomedical Research Centres.Disclosures SH has been an advisor to Abbott Vascular and holds share options in Octiocor. LR received research grants to the institution by Abbott, Biotronik, Boston Scientific, Infraredx, Sanofi, Regeneron and consultation/speaker fees by Abbott, Biotronik, Gentuity, Medtronic, Novo Nordisc, Occlutec. FP is a consultant for Abbott, Amgen, and Novo Nordisk; and has received speaker fees/honoraria from Sanofi. MR and MB are founders of Octiocor Ltd.Introduction Intracoronary optical coherence tomography (OCT) can identify changes following drugs or devices and high-risk plaques causing major adverse cardiovascular events (MACE), but detailed OCT analysis requires time-consuming expert clinician or core laboratory analysis, whilst artifacts and limited sampling impair reproducibility. Assistive technologies such as artificial intelligence (AI)-based analysis may aid both interpretation and patient management, but many studies exhibit methodological, dataset, and reporting deficiencies that limit their robustness for clinical application.Methods Overall, we analysed 366 OCT pullbacks from 297 patients (58,840 OCT frames) with coronary artery disease to develop and test AutoOCT; a modular deep learning AI-based diagnostic aid. The software was designed to correct poor quality or artifact-containing OCT images, identify tissue/plaque composition, classify plaque types, measure multiple parameters including lumen area, lipid and calcium arcs, and fibrous cap thickness (FCT), and output segmented images and clinically useful parameters. 36,212 unselected frames (127 whole pullbacks, 106 patients) were used to train the system. Validation of tissue and plaque classification used ex-vivo OCT pullbacks with co-registered histopathology, while external validation used core laboratory analysis of the high-intensity statin (HIS) IBIS-4 (83 baseline/follow-up patients) and natural history CLIMA (62 patients) studies respectively to determine if AI-based systems can detect plaque progression or regression, changes in plaque composition and higher-risk features with drugs, and future MACE.Results AutoOCT recovered images containing common artifacts and had a plaque classification accuracy of 83% vs. histology, equivalent to expert clinician readers, and replicated core laboratory plaque composition changes after HIS, including reduced lesion lipid arc (13.3° vs. 12.5°) and increased minimum fibrous cap thickness (FCT, 18.9µm vs. 24.4µm). AutoOCT also identified high-risk plaque features leading to MACE including minimal lumen area <3.5 mm 2, Lipid arc >180°, and FCT <75µm, comparable to the CLIMA core laboratory.Conclusions Artifact-corrected, AI-based analysis of intracoronary OCT allows rapid analysis of all available data, identifies tissue and plaque types, and measures features of plaque stabilisation and high-risk plaques. AI-based OCT analysis may assist real-time clinical interpretation and augment clinician or core laboratory analysis for trials of drug/device efficacy and identifying high-risk lesions.Introduction Intracoronary optical coherence tomography (OCT) can identify changes following drugs 1 and high-risk plaque features causing major adverse cardiovascular events (MACE).2–5 However, real-world OCT pullbacks contain hundreds of images and tens-of-thousands of candidate measurements/artery. Consequently, detailed analysis requires time-consuming manual frame selection and measurement in specialised core laboratories and is limited by individual interpretation,6–9whilst artifacts and limited sampling impair reproducibility.10 Assistive technologies such as artificial intelligence (AI)-based analysis show promise, but can have methodological, dataset, and reporting deficiencies, and many models are not sufficiently robust for clinical application.11 We determined if AI-based OCT analysis (AutoOCT) can rapidly process, optimise and analyse OCT images, and identify plaque composition changes that predict drug success/failure and high-risk plaques.Methods Study PopulationModel development used 36,212 OCT frames (127 complete OCT pullbacks,106 unselected patients) from three UK cardiothoracic centres (Cambridge, Swansea, and Chertsey). All pullbacks were included with no exclusion criteria. Histopathological validation used a co-registered OCT and histology dataset.12 External validation used 83 patients from the IBIS-4 OCT arm (NCT00962416)1 and 62 patients from the CLIMA study (NCT02883088).5 AutoOCT Development and TrainingAutoOCT was designed using a DeepLabv3+ architecture, in Python (3.8), with modules to detect the guide catheter or stents, segment artery or imaging components, correct common artifacts, and then segment and measure plaque components. The model was trained with annotated frames in axial cross-sections following accepted plaque definitions,13 and using hybrid dice and cross entropy loss. Data were randomly divided into training, testing and validation sets in a 14:1:1 patient level ratio. A novel OCT image optimisation technique based on histogram matching was used to remove artifacts and improve segmentation accuracy in complex plaque morphologies (figure 1). An additional plaque classification module utilising an EfficientNet architecture was developed from IBIS-4 OCT frames and divided with a patient-level stratification into training (7,904 frames), validation (2,878 frames), and testing (3,246 frames) sets. For classification, we characterised vessel segments as: 1) Low-risk (normal, AIT, PIT), 2) Higher-risk (fibrocalcific, ThCFA and TCFA), with more detailed classification based on measurement of plaque components (e.g. FCTmin).Results Model PerformanceAutoOCT performed well on testing data for whole pullback components (Dice: Lumen 0.99, EEL 0.99, Guidewire shadow 0.96, Lipid 0.84, Calcium 0.85, Fibrous cap 0.80). AutoOCT analysis of a full pullback comprising 271–540 frames takes ~180–300s.Validation Against HistopathologyAutoOCT was validated using ex-vivo OCT pullbacks with matched histopathology.12 128 unique OCT frames were co-registered, and lesions classified histologically by a cardiovascular pathologist (MG).AutoOCT was able to describe histologically defined plaque-types with a similar accuracy to an expert OCT reader (SH), with diagnostic accuracy of 70–91% for different lesions, and 78.1% for TCFA, and non-inferior (p=<0.025 for all plaque-types)(figure 2)(table 1).External Validation - Drug EfficacyWe analysed serial OCT imaging from all 83 patients (153 arteries) of the IBIS-4 OCT sub-study.1 AutoOCT lipid arc measurements correlated well with core laboratory measurements (ICCa 0.75, p=<0.001, average difference (18.3±58.8°, p=<0.001)) and 93.6% (1140/1218) measurements within 95% CI (figure 3A). Whilst AutoOCT FCTmin ICCa 0.66, p=<0.001), average difference (3.1±94.6µm, p=0.241)) and 93.7% (1297/1384) measurements within the 95% CI (figure 3B). Whole-vessel AutoOCT FCTmin and lipid arc showed a similar change to core laboratory analysis (FCT 62.9±28.4μm to 81.8±33.4μm, p=<0.001 vs. 64.88±19.89μm to 87.88±38.08μm, p=0.008; lipid arc 63.1±21.7° to 49.8±20.3°, p=<0.001 vs. 55.94±31.04° to 43.46±3.48°, p=0.013).External Validation - High-Risk Plaque FeaturesThe CLIMA study5 showed that MLA <3.5 mm2, FCT <75µm, and lipid arc >180° were associated with 1-year MACE. We studied 62 participants, comprising 31 MACE and 31 control cases. Although the sensitivity and specificity of each OCT criteria to predict MACE varied, the PPV, NPV and diagnostic accuracy of each variable measured by AutoOCT and the core laboratory were similar (table 2).Discussion We designed and tested an AI-based image analysis system for OCT. Our key findings are (a) AutoOCT could recover images containing common artifacts; (b) plaque classification correlated well with histology; (c) AutoOCT accurately measured FCTmin and lipid arc compared to the core laboratory; (d) AutoOCT replicated core laboratory findings consistent with plaque stabilisation and plaque vulnerability.Many deep learning models are trained and tested with datasets containing limited disease diversity, and with selected frames that exclude common artifacts which may not represent real-world algorithm performance.11 AutoOCT was trained with whole unselected pullbacks (average 285 frames/patient), which is crucial for generalisability, and used pre-processing to mitigate artifacts, allowing analysis of all available data. Additionally, AutoOCT was validated against histopathology as well as core laboratory analyses of individual frames. Whilst improvements are still being made, the current algorithm replicated core laboratory performance.AI-based OCT analysis may aid drug or device development, and trial design. Increased FCT, reduced lipid arc, and TCFA regression can represent a ‘signature’ of a therapy likely to reduce MACE. AutoOCT measurements showed high accuracy, identifying features of drug efficacy. Further, studies identifying plaque vulnerability require large patient numbers often studied for 3–5 years. Analysis is labour-intensive, time-consuming, and requires expert interpretation. AutoOCT had a frame-level accuracy to detect TCFA of 78.1% ex-vivo representing non-inferior performance compared to an expert-reader. While AI-based OCT analysis may not replace core laboratories, whole vessel analysis in minutes/pullback may greatly speed up investigation.Abstract A Figure 1Results of segmentationExamples of model segmentation in frames containing artifacts. From left to right, OCT image, manual annotations, AutoOCT results before, and after optimisation, respectively. (1) Gas bubble artifact; (2) Macrophage dots; (3) Plaque rupture. Arrows denote artifacts, outlined areas denote segmentation errors.Abstract A Figure 2AutoOCT performance in high-risk lesionsLeft to right, Plaque components for higher-risk plaque-types measured on histology sections, co-registered OCT frames by AutoOCT, and expert reader, with plaque classification attention mapAbstract A Figure 3AutoOCT validation of drug effects(A-B) Bland-Altman plots of mean (x axis) and difference (y axis) with histograms of mean (top) and difference (right) for measurements of lipid arc (n=1218)(A) and FCTmin (n=1384)(B). (C-D) Example fibroatheroma lesions that show regression of TCFA (C) or progression of ThCFA (D) after statin therapyAbstract A Table 1Accuracy of AutoOCT and expert reader plaque classification Histological Classification AutoOCT Low Risk ThCFA TCFA Fibrocalcific Sensitivity, (%) 72.6% 70.2% 27.3% 12.5% Specificity, (%) 88.3% 70.4% 88.7% 96.7% PPV, (%) 80.4% 57.9% 33.4% 20.1% NPV, (%) 83.0% 80.3% 85.4% 94.3% Diagnostic Accuracy, (%) 82.0% (0.012) 70.3% (0.001) 78.1% (0.001) 91.4% (<0.001) Expert OCT Reader Sensitivity, (%) 72.7% 53.2% 31.8% 50.0% Specificity, (%) 87.8% 72.8% 84.0% 91.7% PPV, (%) 76.9% 53.2% 29.2% 28.6% NPV, (%) 85.2% 72.9% 85.6% 96.5% Diagnostic Accuracy, (%) 82.4% 65.6% 75.0% 89.1% P values (brackets) demonstrate non-inferiority between AutoOCT and Expert reader.Abstract A Table 2Accuracy of AutoOCT to detect higher-risk plaque features Sensitivity (%) Specificity (%) PPV (%) NPV (%) Core Laboratory Minimum lumen area <3.5 mm2 27.7 86.0 6.8 96.9 Minimum fibrous cap thickness <75µm 40.6 83.9 8.6 97.4 Maximum lipid arc extension >180° 46.9 65.6 4.8 97.1 AutoOCT Minimum lumen area <3.5 mm2 36.7 80.0 6.5 97.1 Minimum fibrous cap thickness <75µm 30.0 86.7 7.9 97.0 Maximum lipid arc extension >180° 56.7 56.7 4.6 97.2 Statement of Contribution Benn Jessney – Data collection and curation, software design, study design, data and statistical analysis, preparation and editing of manuscriptXu Chen – software design and supportSophie Gu – data collectionYuan Huang – statistical analysisMartin Goddard – preparation and analysis of post-mortem materialAdam Brown – data collectionDaniel Obaid – data collectionMichael Mahmoudi – data collectionHector M Garcia Garcia – data collectionStephen P Hoole – data collectionCharis Costopoulos – data collectionLorenz Räber – data collectionFrancesco Prati – data collectionCarola-Bibiane Schönlieb – data collectionMichael Roberts – software design and supportMartin Bennett – supervision and mentorship, editing of manuscript, fundingReferences Raber L, Koskinas KC, Yamaji K, et al. Changes in coronary plaque composition in patients with acute myocardial infarction treated with high-intensity statin therapy (IBIS-4): a serial optical coherence tomography study. JACC Cardiovasc Imaging. 2019;12:1518–1528.Mizukoshi M, Kubo T, Takarada S, et al. Coronary superficial and spotty calcium deposits in culprit coronary lesions of acute coronary syndrome as determined by optical coherence tomography. Am J Cardiol. 2013;112:34–40.Qin Z, Cao M, Xi X, et al. Cholesterol crystals in non-culprit plaques of STEMI patients: a 3-vessel OCT study. Int J Cardiol. 2022;364:162–168.Taruya A, Tanaka A, Nishiguchi T, et al. Vasa vasorum restructuring in human atherosclerotic plaque vulnerability: a clinical optical coherence tomography study. J Am Coll Cardiol. 2015;65:2469–77.Prati F, Romagnoli E, Gatto L, et al. Relationship between coronary plaque morphology of the left anterior descending artery and 12 months clinical outcome: the CLIMA study. Eur Heart J. 2020;41:383–391.Brown AJ, Jaworski C, Corrigan JP, et al. Optical coherence tomography imaging of coronary atherosclerosis is affected by intraobserver and interobserver variability. J Cardiovasc Med (Hagerstown). 2016;17:368–73.Gerbaud E, Weisz G, Tanaka A, et al. Multi-laboratory inter-institute reproducibility study of IVOCT and IVUS assessments using published consensus document definitions. Eur Heart J Cardiovasc Imaging. 2016;17:756–64.Martin WG, McNaughton E, Bambrough PB, West NEJ, Hoole SP. Interobserver variability between expert, experienced and novice operator affects interpretation of optical coherence tomography & 20 MHz intravascular ultrasound imaging. Cardiovasc Revasc Med. 2022.Gruslova AB, Singh S, Hoyt T, et al. Accuracy of OCT core labs in identifying vulnerable plaque. JACC Cardiovasc Imaging. 2024;17:448–450.Jessney B, Chen X, Gu S, et al. Correcting common OCT artifacts enhances plaque classification and identification of higher-risk plaque features. Cardiovasc Revasc Med. 2024.Biccire FG, Mannhart D, Kakizaki R, Windecker S, Raber L, Siontis GCM. Automatic assessment of atherosclerotic plaque features by intracoronary imaging: a scoping review. Front Cardiovasc Med. 2024;11:1332925.Brown AJ, Obaid DR, Costopoulos C, et al. Direct comparison of virtual-histology intravascular ultrasound and optical coherence tomography imaging for identification of thin-cap fibroatheroma. Circ Cardiovasc Imaging. 2015;8:e003487.Tearney GJ, Regar E, Akasaka T, et al. Consensus standards for acquisition, measurement, and reporting of intravascular optical coherence tomography studies: a report from the international working group for intravascular optical coherence tomography standardization and validation. J Am Coll Cardiol. 2012;59:1058–72.