Deep Learning Based Approach for Assessment of Primary Sjögren’s Syndrome from Salivary Gland Ultrasonography Images

Abstract

Salivary gland ultrasonography (SGUS) has shown a good potential for diagnosing Primary Sjögren’s syndrome (pSS). However, existing scoring procedures (based on the manual analysis and grading of images) need further improvements before being established as standardized diagnostic tools. In this study we developed a deep learning based approach for fast and accurate segmentation of salivary glands extended with the scoring of pSS. Total 471 SGUS images were annotated in terms of semantic segmentation and de Vita scoring system. The dataset has been augmented using standard technique (rotation, flip, random crop) and used for training of a deep learning method for segmentation and classification. Our model achieved 0.935 intersection over union (IoU) for segmentation of salivary glands and 0.854 accuracy for classification of pSS stage on validation images. Here, we give an overview of these achievements and show the results.