Background Primary Sjögren's syndrome (pSS) is an autoimmune inflammatory disease predominantly affecting the salivary and lacrimal glands. The main symptoms are dryness of the mouth and eyes. The diagnosis of pSS is based on 6 items according to the American-European consensus group (AECG) classification criteria. Interest regarding ultrasonography (US) as a diagnostic tool for pSS has increased over the last years. However, the applied scoring-systems vary and as of yet international consensus on how to perform the evaluation are lacking. Consequently, the examination and evaluation depends on the examiners skill and experience. In both clinical work and a scientific setting there is a need to use objective methods with as little inter- and intra-examiner variation as possible.
Objectives The aim of this study was to develop a software able to analyze changes in digitally stored US images of the major salivary glands.
Methods Digitally stored US images of glandula parotis (n=94) were blindly evaluated and scored as “normal” or “SS-like” by three independent clinical researchers. At least 2/3 evaluations had to be in agreement to classify the image. All images were from patients fulfilling the AECG criteria. Images had been obtained by six clinical investigators using similar protocols on different US machines. For the analysis, images were divided into databases of SS-like changes and normal-appearing morphology. The image classification and analysis was performed using the ScatNet (v.02) algorithm (Ref: http://www.di.ens.fr/data/software/)for MatLab, which is an algorithm for advanced pattern recognition. Images from the databases were randomly selected to be used as either “training” images or “test” images. Each database of “training” images was analysed, and features of pathological and non-pathological morphology were identified by the software. For the software test, the algorithm analyses which of the databases of training images are most similar to the test images and decides in which group the test images belong. This selection was then compared to the manual scoring.
Results Out of the 94 images evaluated, 40 were classified as normal-appearing and 54 as corresponding to SS-like pathological changes. In the preliminary simulations we have used the following training: test ratios 84:10 (90%), 70:24 (75%) and 47:47 (50%), to respectively train the classification algorithm, and then test the algorithm. The best result with 9/10 correctly classified (92% accuracy) was obtained using 90% of the images to train the software and 10% to test the software. Using 75% or 50% of the images to train the software, the accuracy was reduced to 21/24 (88%) and 25/36 (78%), respectively.
Conclusions Preliminary results indicate that the success rate of the algorithm is closely dependent on the number of images used to train the algorithm. The results are promising and indicate possible clinical use in the evaluation of SGUS images. We will further focus on expanding the image database to achieve more precise results.
Disclosure of Interest None declared