P166 How well does deep learning differentiate between optoacoustic and optical nailfold capillaroscopy images from patients with systemic sclerosis versus those from healthy controls

Abstract Background Nailfold capillaroscopy offers a non-invasive route to observation of systemic sclerosis (SSc)-related microvascular changes and is used routinely for inspection of the capillaries at the finger nailfold. The characteristic changes in capillary structure (increased capillary width, decreased capillary density and abnormal angiogenesis) are included in the 2013 classification criteria for SSc. Optoacoustic mesoscopy is a combination of optical and ultrasound imaging enabling a 3D perspective of capillaries at a similar resolution to commercial nailfold capillaroscopy systems. We have previously reported that quantitative measures of vascular volume and density extracted from optoacoustic images differ significantly between patients with SSc and healthy controls. The aim of this study was to determine whether an artificial neural network (deep machine learning) could correctly differentiate between images from patients with SSc and healthy controls. Methods Optoacoustic (3D, iThera, Germany) and ‘standard’ capillaroscopy images (2D, Optillia, Sweden) of the right and left ring finger nailfolds were acquired. Images were taken at the centre of the nailfold. Acquisition of the same capillaries was difficult in some cases. 2D, greyscale, maximum intensity projections were created from the 3D optoacoustic images. Capillaroscopy images were downsized to match the optoacoustic image resolution. For data augmentation purposes each image, from both the optoacoustic and capillaroscopy data sets, was sliced into multiple overlapping image sections of fixed size. Transfer learning was used to train the model on 'disease' classification (SSc vs control). The pre-trained neural networks learn general image features and subsequently, are fine-tuned on the image data to classify based on the previously learned features. Results Twenty four patients with SSc (19% female; median age 65 IQR [57-69]; duration of Raynaud’s phenomenon 18 [12-28] years; time since onset of first non-Raynaud’s feature 11 (5-18) years) and 19 controls (17% female; age 15 [39-55] years) took part in the study. Fifty random data splits were used to validate the model and showed an average classification accuracy of 0.81 ± 0.15, with an area under the ROC curve of 0.88 ± 0.13 for optoacoustic data. The classification specificity and sensitivity were 0.84 ± 0.22 and 0.77 ± 0.21 respectively for optoacoustic mesoscopy. Performing the same task on capillaroscopy images, achieved an average accuracy of 0.86 ± 0.12 (AUC: 0.92 ± 0.09). Conclusion Deep learning is able to achieve excellent differentiation between images from patients with SSc and controls for both optoacoustic and standard capillaroscopy. Limitations of the study include the relatively small participant numbers and direct comparison of the same capillaries not always being possible. Optoacoustic mesoscopy offers huge potential to increase our understanding of the microvasculature in SSc. Disclosures S. Nitkunanantharajah None. K. Haedicke Corporate appointments; KH is an employee of iThera. T.L. Moore None. J.B. Manning None. G. Dinsdale None. M. Berks None. D. Jüstel None. V. Ntziachristos None. C. Taylor None. M. Dickinson None. A. Herrick None. A.K. Murray None.