3812
View Abstract / View Presentation
Deep Learning based spine labeling with three-plane 2D localizers without vertebrae segmentation
Dattesh Dayanand Shanbhag1, Arathi Sreekumari1, Soumya Ghose2, Chitresh Bhushan2, and Uday Patil3
1GE Healthcare, Bangalore, India, 2GE Global Research, Niskayuna, NY, United States, 3General Electric Company, Bangalore, India
Vertebrae labeling on standard 2D tri-planar MR localizer images is demonstrated without explicit vertebrae segmentation . For lumbar spine MR exam, we report labeling accuracy of 92%, 98% and 96% for T12, L4 and S1 vertebrae on localizer images. 
Figure 4. Sample examples of Spine labeling on 2D localizers for four different test subjects. DL predictions are shown as Red square for T12, green square for L4 and blue square for S1 respectively. A and B : Examples of cases with error much less than acceptable criteria ( < 8 mm error). C: Example with large L4 error and D: Example with large S1 error. Notice that error in one vertebra labelling in C and D doesn’t not affect the accuracy of other labels.
Figure 3. Centroid error for 50 test cases for T12, L4 and S1 labeling. Cutoff for acceptable performance ( 8 mm ) is shown in dotted lines. Most of the errors are within the threshold limits. Overall, accuracy for T12, L4 and S1 is 92%, 98% and 96% respectively. Arrows indicate cases C and D shown in Figure 4 for L4 and S1 error respectively.