Somatostatin receptor PET/MR imaging of large vessel inflammation in active compared with inactive vasculitis and atherosclerosis

Background

Use of 18F-FDG PET in large vessel vasculitis (LVV) is limited by non-specific uptake due to arterial remodelling and/or atherosclerosis leading to diagnostic uncertainty.

Purpose

To investigate somatostatin receptor 2 (SST2) as a novel inflammation-specific PET imaging target in LVV.

Methods

In a prospective observational cohort study, we tested the ability of PET/MRI using two somatostatin receptor tracers (68Ga-DOTATATE and 18F-FET-βAG-TOCA) to differentiate active from inactive LVV, and aortic atherosclerosis in patients with recent myocardial infarction. Ex vivo mapping of the imaging target was performed using immunofluorescence microscopy, imaging mass cytometry, and bulk, single-cell and single-nuclei RNA sequencing of temporal artery biopsies from LVV patients.

Results

Sixty-one participants were included (LVV, n=27; myocardial infarction ≤2 weeks, n=25; control subjects with an oncological indication for imaging, n=9). LVV patients (mean age 58 [SD 16] years; 78% female; 63% active or grumbling disease) had giant cell arteritis (n=13), Takayasu arteritis (n=13), or unspecified LVV (n=1). Baseline index vessel SST2 PET maximum tissue-to-blood ratio (TBRmax) was 61.8% (95% CI 31.5–99.0%, p<0.0001) higher in patients with active/grumbling LVV than inactive LVV, and 34.6% (95% CI 15.1–57.6%, p=0.0002) higher than recent myocardial infarction (Fig. 1a–c; arrow: PET signal; arrowhead: aortic thickening; asterisk: aortic atherosclerosis), with good diagnostic accuracy (AUC ≥0.86, p<0.001 for both). None of the control subjects without LVV or MI had increased arterial SST2 PET signal (Fig. 1d).

Mean aortic TBRmax was strongly correlated with Indian Takayasu Clinical Activity Score (r=0.82 [95% CI 0.46–0.95], p=0.001) and maximum wall thickness on MRI (r=0.68 [95% CI 0.31–0.87], p=0.002). SST2 PET/MRI was generally consistent with 18F-FDG PET/CT in LVV patients with contemporaneous scans (Fig. 1a, b), but with very low background signal in the brain and heart allowing for unimpeded assessment of nearby coronary, myocardial, and intracranial artery involvement. On follow-up imaging after a mean 9.3 (SD 3.2) months, clinically effective treatment for LVV was associated with a 0.49 ±SEM 0.24 (p=0.04; 22.3%) reduction in SST2 PET TBRmax, with good scan-scan repeatability in inactive LVV patients with no change in treatment (ICC 0.86, 95% CI 0.04–0.99).

SST2 localised to macrophages, pericytes, and perivascular adipocytes in inflamed arterial specimens (Fig. 2; a: H&E; b: imaging mass cytometry; arrow: SST2/CD68 co-staining). SSTR2-expressing macrophages co-expressed pro-inflammatory markers (S100A8, S100A9). Specific SST2 radioligand binding was confirmed by autoradiography in LVV specimens.

Conclusion

This is the first study to examine SST2 PET/MRI in LVV and to provide histological and gene expression data for validation. Here we show this novel approach holds major promise for diagnosis and therapeutic monitoring in LVV.

Funding Acknowledgement

Type of funding sources: Foundation. Main funding source(s): Wellcome Trust; Imperial NIHR Biomedical Research Centre

875 Using Stress Testing to Identify Vulnerabilities in Artificial Intelligence Models for the Identification of Culprit Carotid Lesions in Cerebrovascular Events

Introduction

Carotid atherosclerosis is a major risk factor for ischaemic stroke, a leading cause of death. Carotid CT angiography (CTA) is commonly performed following a stroke or transient ischaemic attack (TIA) to help guide patient management in secondary prevention of stroke. Deep learning algorithms can help extract greater information from scans.

Method

The dataset comprised CTA scans from 40 culprit and 40 non-culprit carotid arteries of patients with recent stroke/TIA, and 40 carotid arteries of asymptomatic patients without previous stroke/TIA. A 3D convolutional neural network was trained to classify carotid artery type. Each input comprised 14 axial CTA carotid patches (centred around the carotid artery) concatenated together to form a 3D volume (capturing ∼3cm of artery). 75% of the dataset was used for training and 25% for internal validation. Following training, computer vision operations were applied to input images to assess their impact on the model’s classification decisions.

Results

The model achieved 100% accuracy on the training set and 67% on the internal validation set. However, after subjecting input images to image operations, vulnerabilities in the deep learning model were revealed, even when using input images from the training set. For example, using a Gaussian blur filter with sigma 1.0 was sufficient to change classification decisions, as was horizontally flipping the image.

Conclusions

Deep learning has exceptional capabilities for learning, however the risk with such high-capacity models is failure to learn relevant features from the data. Stress testing provides a viable method to further evaluate deep learning models before clinical deployment.

Radiomics applied to carotid CT angiograms can identify significant differences between culprit and non-culprit lesions in patients with stroke and transient ischaemic attack

Background

Carotid artery atherosclerosis is an important cause of ischaemic stroke. In oncology, textural analysis (“radiomics”) of computed tomography (CT) images can predict the metastatic potential and prognosis of several types of malignant tumours. We investigated whether this quantitative approach could be applied in carotid artery disease.

Purpose

(1) To evaluate the feasibility of computed tomography angiography (CTA) texture analysis in differentiating symptomatic from asymptomatic patients. (2) To investigate whether CTA carotid texture analysis can identify culprit lesions in patients with stroke and transient ischaemic attack (TIA).

Methods

Carotid CTAs of consented research subjects were included in the study. Symptomatic patients had confirmed carotid artery-related ischaemic stroke or TIA in the 7 days before CTA imaging. Asymptomatic (ASX) patients had no prior stroke/TIA. Both TexRAD, a research texture analysis software, and PyRadiomics, a Python package for radiomics studies, were used to extract 99 first-order and higher-order texture features from regions-of-interest (ROI) drawn around the outer wall of the carotid artery. Single-slice analysis compared the carotid bifurcations of symptomatic and asymptomatic patients, and of culprit (CC) and non-culprit (NC) arteries in symptomatic patients. Multi-slice analysis was conducted using a 3D volume defined by ROIs drawn on 14 consecutive CT slices of 3mm thickness, covering 3cm of carotid artery. The Mann-Whitney U test was used for inter-subject comparisons (ASX vs CC; ASX vs NC) and the Wilcoxon signed-rank test was used for intra-subject comparisons (CC vs NC). A p value <0.0005 was deemed statistically significant after Bonferroni correction for multiple comparisons. Non-normally distributed variables are reported as median (interquartile range).

Results

The dataset comprised 82 carotid arteries from 41 symptomatic patients (41 culprit; 41 non-culprit) and 50 carotid arteries from 25 asymptomatic patients. Single-slice analysis revealed greater homogeneity in asymptomatic carotids versus symptomatic culprit carotids (Uniformity: ASX 0.11 (0.05); CC 0.08 (0.05), p<0.0005) and non-culprit carotids (NC 0.08 (0.18), p<0.0005). In multi-slice analysis, culprit and non-culprit carotid arteries displayed greater heterogeneity than asymptomatic carotids (GLSZM zone entropy: CC 6.57 (0.59); NC 6.76 (0.65); ASX 6.21 (0.32), p<0.0005). Multi-slice analysis of symptomatic culprit versus non-culprit carotids revealed greater heterogeneity in culprit carotids than non-culprit carotids (GLRLM run entropy CC 6.57 (0.59); NC 5.05 (0.70), p<0.0001).

Conclusion

Textural analysis of carotid CTAs reveal significant differences between symptomatic and asymptomatic patients and between culprit and non-culprit carotid arteries within symptomatic patients. This approach could be used to identify patients at high risk of further stroke for aggressive medical therapy and surveillance.

Funding Acknowledgement

Type of funding source: Public grant(s) – National budget only. Main funding source(s): EPVL is undertaking a PhD funded by the Cambridge School of Clinical Medicine and the Medical Research Council's Doctoral Training Partnership

Pericoronary adipose tissue density is associated with clinical disease activity in Takayasu arteritis and coronary arterial inflammation measured by 68Ga-DOTATATE PET in atherosclerosis

Background

Coronary artery disease (CAD) is an under-recognized complication of intense arterial inflammation in Takayasu arteritis (TAK). While pericoronary adipose tissue (PCAT) density is associated with arterial inflammation in CAD patients, this relationship has not previously been studied in TAK patients, nor directly compared with coronary arterial inflammation measured by 68Ga-DOTATATE positron emission tomography (PET).

Purpose

To compare PCAT density with clinical, biochemical and molecular imaging markers of inflammation in TAK and CAD patients.

Methods

PCAT density was quantified from computed tomography coronary angiography (CTCA) around each of the 17 coronary segments in patients with: (1) TAK and CAD, (2) atherosclerotic CAD, and (3) age and gender-matched healthy controls, using semi-automated software (Autoplaque). In TAK patients, PCAT density was compared to the Indian Takayasu Clinical Activity Score (ITAS) and high-sensitivity C-reactive protein (CRP). In CAD patients, PCAT density was compared to local arterial inflammation measured by coronary motion-frozen 68Ga-DOTATATE PET using image registration software (FusionQuant), and systemic (aortic) inflammation using 18F-fluorodeoxyglucose (FDG) PET. Data was acquired either during routine clinical care or prior research that established 68Ga-DOTATATE as an experimental marker of arterial inflammation that binds macrophage somatostatin receptor-2 in atherosclerotic plaques (NCT02021188).

Results

60 patients were included (TAK, n=20; CAD, n=20; healthy, n=20). Non-calcified plaque burden (TAK: 95.2%; CAD: 90.4%, p<0.0001) and CRP (TAK: 25.2 ±SD 16.1 mg/L; CAD: 2.5 ±SD 1.7 mg/L, p=0.04) were greater in TAK than CAD patients.

PCAT density varied significantly among the three groups (median [IQR] TAK: −72.9 [−81.2 to -66.1] Hounsfield unit [HU]; CAD: −79.9 [−88.0 to −72.2]; healthy: −83.8 [−90.1 to −75.8] HU, p<0.0001). Figure: box-plot showing the distribution of PCAT values by group, with corresponding representative multiplanar reconstructed and cross-sectional CTCA images with surrounding PCAT density displayed by color table in left anterior descending arteries.

PCAT density was significantly associated with ITAS (r=0.61, p=0.004) and CRP (r=0.43, p=0.03) in TAK patients, and coronary 68Ga-DOTATATE maximum tissue-to-blood ratio (r=0.31, p<0.001) in CAD patients. PCAT density was not associated with aortic 18F-FDG uptake in CAD patients, nor subcutaneous (pre-sternal) adipose tissue density in either disease group. No significant patient-level confounders were identified using linear mixed-effects regression modelling.

Conclusion

PCAT density measured by CTCA is greater in TAK than CAD patients, and is associated with clinical and biochemical markers of disease activity in TAK, and coronary arterial inflammation measured by 68Ga-DOTATATE PET in CAD. PCAT could be a useful, easy to measure marker of coronary inflammation and disease activity in both TAK and CAD.

PCAT density is greater in TAK than CAD

Funding Acknowledgement

Type of funding source: Foundation. Main funding source(s): Wellcome Trust

Contrast CT classification of asymptomatic and symptomatic carotids in stroke and transient ischaemic attack with deep learning and interpretability

Background

Convolutional neural networks (CNNs), part of deep learning, are used widely for computer vision tasks and in some medical domains, such as mammography interpretation. The application of deep learning to carotid artery imaging is scarce. We investigated the ability of deep learning to correctly classify contrast CT images of the carotid arteries without the need for prior feature selection.

Purpose

(1) To assess the ability of deep learning to differentiate symptomatic patients (had prior stroke or transient ischaemic attack [TIA]) from asymptomatic patients (no prior stroke/TIA) using contrast CT scans alone. (2) To investigate whether deep learning can further discriminate between culprit and non-culprit carotid arteries in symptomatic patients. (3) To assess the interpretability of the deep learning models.

Methods

Carotid contrast CT scans of consented research subjects were included in the study. Symptomatic patients had confirmed carotid artery-related ischaemic stroke or TIA in the 7 days before CT imaging, and asymptomatic patients had no prior cerebrovascular events. The dataset comprised 1148 axial symptomatic slices (covering a 3cm area of each carotid artery in 41 patients; 41 culprit and 41 non-culprit carotids) and 700 asymptomatic slices (from the bilateral carotid arteries of 25 patients). The dataset was split such that 75% was used for training and 25% for testing. A 30x30 bounding box was used to create patches of the carotid arteries from these axial slices for use as input to the CNN, a modified VGG16 architecture initialised with ImageNet weights to leverage transfer learning (the application of a model trained in one domain to a different domain) implemented in Python. Data augmentation was applied to the training set and the model was trained for 100 epochs using a cyclic learning rate, the RMSProp optimizer and binary cross-entropy loss. Class activation heatmaps were generated using the GradCAM method to highlight the areas of the image that were most important to the model for making its classification decision.

Results

The deep learning model was 92% accurate in correctly identifying carotid arteries from symptomatic patients versus those from asymptomatic patients. Discriminating between culprit versus non-culprit carotid arteries in symptomatic patients alone was 71% accurate. The class activation heatmaps demonstrated how the model learnt to localise the carotid artery within the image patch, and to ignore the arterial lumen when making its classification decision.

Conclusions

Deep learning can be used to differentiate between symptomatic and asymptomatic carotid CT scans from stroke/TIA subjects without the need for prior feature engineering. The model learns to identify relevant features in the image that predict the patients' symptom state. If further validated, this approach could be used to identify high-risk patients for intensive medical therapy.

Funding Acknowledgement

Type of funding source: Public grant(s) – National budget only. Main funding source(s): EPVL is undertaking a PhD funded by the Cambridge School of Clinical Medicine and the Medical Research Council's Doctoral Training Partnership