Erik B Dam PhD, Presenter: Employee, SYNARC Inc

Joselene Marques MSc, Abstract Co-Author: Grant, Synarc, Inc

Osteoarthritis involves changes in bone shape and structure, articular cartilage shape and structure, meniscus, synovium, and ligaments. Magnetic resonance imaging (MRI) allows more comprehensive inspection of bone and cartilage in multiple compartments compared to traditional radiography. For MRI to be available for large clinical trials with thousand(s) participants and several visits, automation of the scoring seems desirable. We evaluated a method for fully automatic segmentation of bone and cartilage from knee MRI. 

MRI were acquired using a Turbo 3D T1 sequence from a 0.18T Esaote scanner (40° FA, TR 50ms, TE 16 ms, scan time 10 minutes, resolution 0.7mm x 0.7mm x 0.8mm). Radiographs were acquired in a load-bearing semiflexed position using the SynaFlex and graded by an experienced radiologist on the Kellgren and Lawrence index (KL).  The study included 145 knees (51 with KL 0, rest evenly KL1-3) from 22-79 years old subjects with 59% female. 31 representative knees were rescanned a week later to allow inter-scan precision validation. The Tibia bone and the medial tibial and femoral cartilages were segmented automatically using a computer-based voxel classification framework. The method was trained on 25 knees and evaluated on the rest. All scans were manually segmented by a radiologist by slice-wise outlining. 

All scans were successfully segmented. Precision was measured as inter-scan coefficient of variation (CV). For the three compartments, CV on volume quantifications were 5.3%, 6.3%, and 10.0% for Tibia, tibial cartilage, and femoral cartilage, respectively. Segmentation accuracy was measured as the Dice volume overlap. Dice were 97%, 81%, and 79%, for Tibia, tibial cartilage, and femoral cartilage, respectively. For comparison, repeated manual radiologist segmentations gave precision of CV 7.3% and 6.8% and accuracy of Dice 86% and 86%, for tibial and femoral cartilages respectively. 

The results demonstrate the feasibility of performing automatic segmentation of bone and cartilage in knee MRI with a performance close to manual segmentation by a trained radiologist. 

For diagnosis of the individual, automatic segmentation and diagnosis may not yet be feasible. However, for large clinical studies, cartilage volume may be effectively quantified by an automatic metho

Dam, E, Marques, J, Automatic Segmentation of Bone and Cartilage from Knee MRI.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11008804.html