Scoring Primary Sjögren's syndrome affected salivary glands ultrasonography images by using deep learning algorithms

Abstract

Salivary gland ultrasonography (SGUS) represents a promising tool for diagnosing Primary Sjögren's syndrome (pSS), which is manifest with abnormalities in salivary glands (SG). In this study, we propose a fully automatic method for scoring SGs in SGUS images, which is the most important step towards SG the pSS diagnosis. A two-centric cohort included 600 images (150 patients) annotated by experienced clinicians. The aim of the study was to assess various deep learning classifiers (MobileNetV2, VGG19, Dense-Net, Squeeze-Net, Inception_v3, and ResNet) for the purpose of the pSS scoring in SGUS. The training was performed using the ADAM optimizer and cross entropy loss function. Top performing algorithms were MobileNetV2, ResNet, and Dense-Net. The assessment showed that deep learning algorithms reached clinicians-level performances in the almost real-time. Considering that, the further work should be regarded towards evaluation on larger and international data sets with the goal to establish SGUS as an effective noninvasive pSS diagnostic tool.