POS1391 DEEP LEARNING-BASED PANNUS LOCALIZATION SOFTWARE IN THE HANDS OF INFLAMMATORY ARTHRITIS USING TIME-INTENSITY CURVE (TIC) SHAPE CLASSIFICATION ON DYNAMIC MRI DATASET

Abstract

Background: Rheumatoid arthritis (RA) is a chronic infammatory disease associated with signifcant functional impairment and disability, linked to infammatory and structural articular and peri-articular damage [1]. Synovitis is a characteristic feature of RA, and is considered an important factor in disease activity and the best predictive marker of joint damage [2]. Therefore, accurate quantifcation of synovitis can play an important role in clinical evaluation and treatment serving as a biomarker. Pixel-by-pixel time-intensity curve (TIC) shape analysis is a dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) technique to help visualize differently shaped TICs [3]. Pixels having TIC shape type 4, which is characterized as early enhancement followed by washout phase, were regarded as synovitis pixel [4]. A deep residual network called ResNet-50 model, a convolutional neural network (CNN) that is 50 layers deep, might effectively classify TIC shapes. Objectives: This study aimed to develop software that can automatically demonstrate the distribution of enhancing synovial pannus of patients with infammatory arthritis on DCE-MRI. Methods: Modifed ResNet-50 was used on MATLAB. Two investigators drew regions of interest (ROIs) of muscle, bone, and synovitis under expert guidance and obtained TICs for each tissue on DCE-MRI of ten rheumatoid patients. Through cross-validation and batch evaluation, we verifed the confusion matrices for the performance of the classifer. Then we identifed the pixels of enhancing synovial pannus by image classifcation from the obtained pixel-by-pixel TIC shape and displayed them in color. The software performance was evaluated using a visual assessment on seven hand joints of one patient. Results: 150 ROIs for muscles, 150 ROIs for bones, and 59 ROIs for synovitis on DCE-MRI of the hand joints were drawn in ten patients and obtained 4049, 3825, and 1041 TIC shape images, respectively. The classifer's accuracy, precision, sensitivity, and specifcity were 99.6%, 99.3%, 98.4%, and 99.7%, respectively. Out of seven joints, four were assessed as good, two as fair, and one as fail. Conclusion: Our classifer showed high accuracy, precision, sensitivity, and specifcity. And through the comparison between manual outlining and the result of software, our software had relatively good performance. This automatic software developed using deep learning with CNN might accurately display the enhancing synovial pannus in RA.