skip to main content
Language:
Search Limited to: Search Limited to: Resource type Show Results with: Show Results with: Search type Index

AB1335 BONE MARROW EDEMA SCORE IN HAND X-RAY FILM BY AI DEEP LEARNING ASSOCIATE WITH MRI BONE EDEMA IN RHEUMATOID ARTHRITIS

Annals of the rheumatic diseases, 2022-06, Vol.81 (Suppl 1), p.1773-1774 [Peer Reviewed Journal]

ISSN: 0003-4967 ;EISSN: 1468-2060 ;DOI: 10.1136/annrheumdis-2022-eular.933

Full text available

Citations Cited by
  • Title:
    AB1335 BONE MARROW EDEMA SCORE IN HAND X-RAY FILM BY AI DEEP LEARNING ASSOCIATE WITH MRI BONE EDEMA IN RHEUMATOID ARTHRITIS
  • Author: Katayama, K. ; Pan, D. ; Oda, M. ; Okubo, T. ; Mori, K.
  • Is Part Of: Annals of the rheumatic diseases, 2022-06, Vol.81 (Suppl 1), p.1773-1774
  • Description: Background Rapid radiographic progression (RRP) was reported to be one of clinical symptom in difficult to treat RA(D2T RA) (1). Eular recommendation for imaging showed BME is strong and independent prognostic factor for bone destruction(2). We reported bone marrow edema (BME) in MRI image was most associated with RRP compared with bone erosion, synovitis in Adalimumab add-on therapy in MTX-IR RA patients(3). To rescue RRP, early detection of BME is important although cost of MRI is expensive and hard to repeat. Objectives To investigate the score of BME in hand X ray film by deep learning between X ray film and MRI BME information can discriminate the differences between BME and non-BME images. Methods In this work, we use a neural network consisted of convolutional layers and fully connected layers to classify X-ray images (Figure 1) In this paper, the output is the socre of BME which ranges from 0 to1(threshold = 0.4). We also used an interpretation technique called the Grad-CAM for visual explanations. Hand MRI (1.5T) were used. Figure 1. The convolutional neural network design. A red block “Conv” means a convolutional block. It contains a 2D convolution layer, a leaky relu activation function, a maxpooling layer and a batch normalization layer. The numbers above each “Conv” block are (kernal size, kernal size, kernal number). A green block “FC” is a fully connection layer. The number above it is (neuron number). After the last Softmax layer, the output becomes the probability of BME which ranges from 0 to 1. Results Regarding data split, 104 images including 79 non-BME images and 25 BME images are used as a hold-out test set. The rest of the images (473 images) are used as training data and validation data. Five fold cross-validation is used for these 473images. For each fold, there are about 378 images including 297 non-BME images and 81 BME images in the training set. There are about 95 images including 74 non-BME images and 21 BME images in the validation set. In order to fully utilize every image and unify the distribution of the training set and the validation set, the ratio of non-BME and BME is controlled to be the same which is about 3.66:1. The five folds showed similar performance on the hold-out test set. AUC is the area under the ROC curve. As the result, AUC which indicates the general performance of this model, ranged from 0.88 to 0.91. The average precision was 63% and the average recall rate was 87%. In this experiment, the initialization seed will greatly influence the final result. For example, AUC can be reduced to 0.73 from 0.89 because of a different initialization seed. It perhaps results from the shortage of data, which can easily make the neural network drop into a local minimum.We also utilized Grad-cam to visualize the result. The result of Grad-cam shows the importance of each part to the final prediction(Figure 2). Figure 2. Result of Grad-cam. Numbers in the parenthesises are the possibilities of BME. The middle case is unexpected because red region is the surrounding of the hand. The left and right cases may indicate the evidence for prediction. Conclusion The preliminary result is much better than a random guess. According to this result, there should be a certain difference between BME and non-BME images. If it’s the characteristic of BME that domains this difference, our classification algorithm will be feasible for BME. Our future work is to justify the evidence of the predictions and improve performance. References [1]Nagy G et al. Eular definition of difficult- to - treat to rheumatoid arthritis. Ann Rheum Dis 2021;80:31-35 [2]Colebatch AN et al. Eular recommendation for the use of imaging of joints in the clinical management of rheumatoid arthritis. Ann Rheum Dis 2013;72: 804-814 [3]Katayama K et al. Bone marrow OEDEMA is more associated with rapid radiographic progression than in synovitis or bone erosion by using low field MRI in bio-naiive rheumatoid arthritis patients treated with adalimumab and methotrexate combination therapy. Ann Rheum Dis 2014, eular meeting SAT0100. Disclosure of Interests None declared
  • Language: English
  • Identifier: ISSN: 0003-4967
    EISSN: 1468-2060
    DOI: 10.1136/annrheumdis-2022-eular.933
  • Source: AUTh Library subscriptions: ProQuest Central
Back to results list
INSPIRE LIBRARY - TON DUC THANG UNIVERSITY (84-028) 37 755 057 Feedback
19 Nguyen Huu Tho St. Dist.7, HCM thuvien@tdtu.edu.vn

Copyright © 2017 INSPIRE LIBRARY - TON DUC THANG UNIVERSITY

Searching Remote Databases, Please Wait