Neural network based algorithm pipeline for automatic detection of erosions and ankylosis of the sacroiliac joints : development and validation using multicentre CT images

Neural network based algorithm pipeline for automatic detection of erosions and ankylosis of the sacroiliac joints: development and validation using multicentre CT images Van Den Berghe T.1, Chen M.2, Babin D.3, Herregods N.1, Huysse W.1, Jaremko J.L.4, Laloo F.1, De Craemer A.-S.5, Carron P.5, Van Den Bosch F.5, Elewaut D.5, Jans L.B.O.1 1Dept. Radiology, Ghent University Hospital, Ghent, Belgium; 2Dept. Radiology, Peking University Shenzhen Hospital, Shenzhen, China; 3Dept. Telecommunication and Informatics, Ghent University, Ghent, Belgium; 4Dept. Radiology, University of Alberta Hospital, Edmonton, Canada; 5Dept. Rheumatology, Ghent University Hospital, Ghent, Belgium Background Axial spondyloarthritis (SpA) typically affects the sacroiliac joints (SIJ) with erosion and ankylosis as structural lesions. Mostly beginning before 40 years old, it is characterized by low back and buttock pain and accounts for 5% of chronic low back pain patients. We aimed to develop and evaluate the diagnostic accuracy of a deep learning based algorithm for the automatic detection and quantification of erosion and ankylosis of the SIJs on CT images. Methods 145 patients (81 female, 64 male, 121 Ghent University Hospital, 24 University of Alberta Hospital, 18-87 years old, mean age 40±13, 84 axial SpA, 15 mechanical back pain, 46 without clear diagnosis but with symptoms suspicious for axial SpA and/or positive family history and/or HLAB27 positivity and/or recurrent anterior uveitis and/or Crohn’s disease) that underwent a CT scan of the SIJs because of clinically suspected sacroiliitis between March 2005 and January 2021 were included retrospectively. Ground truth segmentation of the SIJs was manually performed and quality-controlled by 3 independent experienced radiologists. Erosions >1 mm and ankylosis >2 mm were manually annotated by 3 independent experienced radiologists blinded for clinical diagnosis. A deep learning based preprocessing, U-Net and convolutional neural network (CNN) pipeline was developed to automatically segment the SIJs and detect structural lesions. Internal in-training cross validation was performed to assess the diagnostic performance of the algorithm on a lesion and patient level. Results Regarding segmentation, a dice similarity coefficient of 0.75±0.03 was obtained. For slice by slice validation group lesion detection, erosions were depicted with an accuracy of 89%, a positive predictive value (PPV) of 89%, a negative predictive value (NPV) of 90%, a sensitivity of 90% and a specificity of 89%. Ankylosis was depicted with an accuracy of 91%, a PPV of 91%, a NPV of 93%, a sensitivity of 93% and a specificity of 91%. For patient validation group lesion detection, erosions were depicted with an accuracy of 74% for a threshold confidence level (TCL) of 98% and a threshold number of windows (TNW) of 15. Ankylosis was depicted with an accuracy of 88% for a TCL of 70% and a TNW of 27. Optimization steps towards reduction of false positives or negatives were performed. Conclusion Erosion and ankylosis in patients with sacroiliitis can be automatically detected on CT images with a high accuracy and in an objective way using a deep learning based neural network pipeline.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.