Ismail Ben Ayed, Presenter: Research, General Electric Company

Seyed-Parsa Hojjat PhD, Abstract Co-Author: Nothing to Disclose

Kumaradevan Punithakumar PhD, Abstract Co-Author: Nothing to Disclose

Gregory J. Garvin MD, Abstract Co-Author: Nothing to Disclose

Precise detection and identification of spine structures (e.g., the vertebrae and discs) facilitate the diagnosis of various spine disorders. For instance, in MRI, spine labeling provides anatomical benchmarks that ease dramatically the evaluation and reporting of frequent disc deformities, e.g., protrusion. In addition to their usefulness in spine diagnosis, such benchmarks yield a patient-specific coordinate system that can be very useful in (i) mapping radiologic reports to the corresponding image segments, (ii) building semantic inspection tools, (iii) guiding image registration, and (iv) providing priors for segmentation, image retrieval, as well as shape and population analysis.  

We propose an efficient (nearly real-time) two-stage spine MRI labeling algorithm, which is applicable to different types of MRI data and acquisition protocols. The first stage aims at roughly detecting vertebra candidates with a novel segmentation technique. The second stage removes false positives and identifies all vertebrae and discs by imposing anatomical constraints. We performed quantitative evaluations over 90 mid-sagittal MRI images of the lumbar spine acquired from 45 subjects. We used both T1- and T2-weighted images for each subject. 990 structures were automatically detected and labeled, and compared to ground-truth annotations by an expert. On the T2-weighted data, we obtained an accuracy of 91:6% for the vertebrae and 89:2% for the discs. On the T1-weighted data, we obtained an accuracy of 90:7% for the vertebrae and 88:1% for the discs. 

Most of the existing algorithms require intensive and time consuming training from a large and manually-labeled data set, with the results often being dependent on the choice of (i) the training set and (ii) the modality and/or acquisition protocol. Our algorithm removes training requirements and, therefore, promises to handle the substantial variations encountered in realistic clinical contexts. 

The proposed algorithm removes the need for training, while being applicable to different types of MRI data and acquisition protocols. Our comprehensive evaluations over T1- and T2-weighted images demonstrated the flexibility of the algorithm. 

http://abstract.rsna.org/uploads/2014/14011365/14011365_e8em.jpg

Ben Ayed, I, Hojjat, S, Punithakumar, K, Garvin, G, Automatic Data Driven Labeling of Lumbar Spine Structures in MRI.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14011365.html