fetching data ... 
Background: Artificial Intelligence (AI) has rapidly established itself as a catalyst for change in many fields, and medicine is not an exception. The emergence of AI in medicine has opened up exciting new possibilities, revolutionizing the way we diagnose and deliver care. One of the most promising advances is in the field of Sjögren’s syndrome. In this context, the use of deep learning techniques applied to glandular ultrasound has played a crucial role in improving the detection and early diagnosis of this disease.
Objectives: The aim of this project is to create a web-based diagnostic application for physicians, with a deep learning model to detect Sjogren’s syndrome.
Methods: The data are from the HarmonicSS project [1]. This project is a source of data for research in primary Sjögren’s syndrome (pSS) and has contributed to the collection of static ultrasound images of the major salivary glands (SGUS), including the parotid glands (PG) and submandibular glands (SMG). In addition, an accompanying Excel spreadsheet contains additional information on each image, OMERACT score was used for the interpretation of SGUS images. Ultrasound images were pre-processed, including image resizing, flipping, rotation, translation, zooming and distortion. After selecting the model, we included other evaluation metrics such as Specificity, False Positive Rate (FPR), False Negative Rate (FNR), Positive Likelihood Ratio (LR+) and Negative Likelihood Ratio (LR-), to better assess its performance in a medical context.
Results: In total, we have collected 225 ultrasound images of the salivary glands from 150 patients with pSS and non-Sjogren’s syndrome. After model construction, we decided to use 5 convolutional layers in our architecture. We found that using 5 convolutional layers allowed our model to learn features of different scales, from simple edges to more complex patterns. At the end, four models have been developed, and one was finally selected as offering the best accuracy.
In the learning phase, the performance of our proposed model has an accuracy of 0.99, a precision of 0.99, sensitivity of 98%, specificity of 99%, (LR+) of 174 and (LR-) of 0.01 for diagnosing Sjogren’s syndrome. In the testing phase, the performance of our model has an accuracy of 0.99, a precision of 0.99, sensitivity of 98% and specificity of 99% for diagnosing Sjogren’s syndrome, (LR+) of 181 and (LR-) of 0.01.
Conclusion: After examining the results of the model explanations, we observed that the extracted superpixels encompass restricted parts of the gland present in the ultrasound. However, they do not capture all of its particular texture. In fact, the gland has a distinct texture and a unique structure, clearly differentiating it from other areas of the image. In addition, there are other superpixels that have been selected from other areas of the image. This observation suggests that the performance of the proposed model could potentially improve and scale up in terms of ability to recognize the overall structure of the gland more completely and accurately. This artificial intelligence model, for interpreting ultrasound images of the salivary glands, is promising for diagnosing Sjogren’s syndrome, while requiring further developments and assessments.
REFERENCES: [1] HarmonicSS,
Acknowledgements: NIL.
Disclosure of Interests: None declared.