An An aut automated ed pr probabilis babilistic ic alg algorit ithm hm for the he de detectio ection n of cen centr tral al vein ein sig ign n in in whit hite-ma matter lesion ons March 1, 2019 Jordan Dworkin
Sati et al., 2016 4 2
Motivation • Can automated methods help detect whether or not a lesion has a central vein? • Why is this important? • A high proportion of central vein lesions seems to be highly specific to MS, compared to other diseases with white matter lesions 4 • Rigorous definitions of what qualifies as a central vein make manual determination a time intensive process, which has lead some to recommend picking a random sample of 3 lesions for judgement • Less rigid definitions would lessen the burden, but exacerbate subjective differences in judgements • Goal • To create a method that automatically detects central vein sign in white matter lesions 3
Proposed method • Why build a rigid, untrained algorithm? • More control over what features it considers important • Potentially more robust to application at different sites • Yields probabilistic estimate of CVS • Points clinicians towards lesions where the model is less confident • Algorithm steps: • Find white-matter lesions • Find veins • Determine centrality of veins within lesions 4
1) Find white-matter lesions 2) Find veins 3) Determine centrality 5
Campbell et al., 2015 2 1) Find white-matter lesions 2) Find veins 3) Determine centrality 6
1) Find white-matter lesions 2) Find veins 3) Determine centrality 7
1) Find white-matter lesions 2) Find veins 3) Determine centrality 8
1) Find white-matter lesions 2) Find veins 3) Determine centrality 9
1) Find white-matter lesions 2) Find veins 3) Determine centrality 10
Permutation procedure • Examining coherence between vesselness, ! "#$ , and centrality, % "#$ , of voxels in a lesion • Subject i , lesion j ’s coherence: • & "# = ∑ $∈* % "#$ ∗ ! "#$ • Want to find the probability of seeing this level of coherence by chance • For , ∈ {1, … , 1 2 } ∗ • & "#2 = ∑ $∈*,4∈* ∗ % "#$ ∗ ! "#4 5 11
Permutation procedure • Examining coherence between vesselness, ! "#$ , and centrality, % "#$ , of voxels in a lesion • Subject i , lesion j ’s coherence: • & "# = ∑ $∈* % "#$ ∗ ! "#$ • Want to find the probability of seeing this level of coherence by chance • For , ∈ {1, … , 1 2 } ∗ • & "#2 = ∑ $∈*,4∈* ∗ % "#$ ∗ ! "#4 5 P cvs = 0.99 12
Performance assessment • Tested algorithm on sample of 39 participants at the University of Vermont 5 • 10 had MS with no white-matter comorbidities • 10 had MS with white-matter comorbidities • 10 had migraine with white-matter abnormalities • 9 were previously incorrectly diagnosed with MS • Wanted to determine whether algorithm would find a higher proportion of central vein lesions in MS patients, and whether it could be diagnostically effective • Sought to test two versions of the algorithm: • 1 – Semi-automated; lesion segmentation is checked and artifacts are removed before CVS detection • 2 – Fully-automated; no QA following segmentation 13
Performance assessment • Fully-automated: Within-patient proportions 1.00 MS of CVS were significantly higher in MS patients No than non-MS patients Yes • Mean CVS proportion in MS patients: 53% 0.75 • Mean CVS proportion in non-MS patients: 34% Fully − automated biomarker Fully–automated marker MS 0.50 0.25 0.00 Migraine Misdiagnosed Multiple sclerosis Multiple sclerosis (with comorbidities) (no comorbidities) Group 14
Performance assessment • Fully-automated: Within-patient proportions 1.00 MS of CVS were significantly higher in MS patients No than non-MS patients Yes • Mean CVS proportion in MS patients: 53% 0.75 • Mean CVS proportion in non-MS patients: 34% Semi − automated marker MS 0.50 • Semi-automated: Differences were more pronounced after removing false-positives from lesion segmentation 0.25 • Mean CVS proportion in MS patients: 57% • Mean CVS proportion in non-MS patients: 27% 0.00 Migraine Misdiagnosed Multiple sclerosis Multiple sclerosis (with comorbidities) (no comorbidities) Group 15
Performance assessment • Semi-automated marker had high sensitivity 1.00 and specificity in distinguishing MS from non- MS participants • Optimal sensitivity/specificity = 0.75/0.95 AUC = 0.87 0.75 • AUC = 0.87 AUC = 0.81 Sensitivity Method • Fully-automated marker also distinguished 0.50 MS from non-MS participants, but tended to have stronger sensitivity than specificity • Optimal sensitivity/specificity = 0.85/0.63 0.25 Method • AUC = 0.81 Fully − auto Semi − auto 0.00 1.00 0.75 0.50 0.25 0.00 Specificity 16
Summary • This study developed an automated method for detecting CVS in white matter lesions • Previous findings that CVS occurs at a higher rate in MS subjects are replicated • Performance provides preliminary evidence for the algorithm’s validity and potential diagnostic utility • Proposed method builds on existing lesion and vein segmentation tools • Allows for the use of site-specific methods, and the continuous implementation of the newest segmentation methods • Could reduce the prevalence of false-positive lesions, and bring the fully-automated results closer to the semi-automated results 17
References 1. Frangi AF, Niessen WJ, Vincken KL, Viergever MA. Multiscale vessel enhancement filtering. In: Wells WM, Colchester A, Delp S, eds. Medical Image Computing and Computer-Assisted Intervention — MICCAI’98 . Vol 1496. Berlin, Heidelberg: Springer Berlin Heidelberg; 1998:130 -137 2. Campbell IC, Coudrillier B, Mensah J, Abel RL, Ethier CR. Automated segmentation of the lamina cribrosa using Frangi’s filter: a novel approach for rapid identification of tissue volume fraction and beam orientation in a trabeculated structure in the eye. J R Soc Interface . 2015;12(104):20141009-20141009 3. Sweeney EM, Shinohara RT, Shea CD, Reich DS, Crainiceanu CM. Automatic Lesion Incidence Estimation and Detection in Multiple Sclerosis Using Multisequence Longitudinal MRI. Am J Neuroradiol . 2013;34(1):68-73 4. Sati P, Oh J, Constable RT, et al. The central vein sign and its clinical evaluation for the diagnosis of multiple sclerosis: a consensus statement from the North American Imaging in Multiple Sclerosis Cooperative. Nat Rev Neurol . 2016;12(12):714-722 5. Solomon AJ, Watts R, Ontaneda D, Absinta M, Sati P, Reich DS. Diagnostic performance of central vein sign for multiple sclerosis with a simplified three- lesion algorithm. Mult Scler Houndmills Basingstoke Engl . August 2017 6. E. M. Sweeney, R. T. Shinohara, N. Shiee, F. J. Mateen, A. A. Chudgar, J. L. Cuzzocreo, P. A. Calabresi, D. L. Pham, D. S. Reich, and C. M. Crainiceanu, “OASIS is Automated Statistical Inference for Segmentation, with applications to multiple sclerosis lesion segmentation in MRI,” NeuroImage Clin. , vol. 2, pp. 402– 413, 2013 7. E. M. Sweeney, R. T. Shinohara, B. E. Dewey, M. K. Schindler, J. Muschelli, D. S. Reich, C. M. Crainiceanu, and A. Eloyan, “Relating multi-sequence longitudinal intensity profiles and clinical covariates in incident multiple sclerosis lesions,” NeuroImage Clin. , vol. 10, pp. 1–17, 2016 18
Acknowledgements • University of Pennsylvania • Taki Shinohara, Ali Valcarcel, and the PennSIVE group • Ipek Oguz and PICSL • Amit Bar-Or, Matthew Schindler, and the Penn MS Division • NINDS • Daniel Reich, Pascal Sati, and Dzung Pham • University of Vermont • Andrew Solomon and Richard Watts • NAIMS Cooperative 19
Recommend
More recommend
