Development Automatic Classification of Vertebral Shape Based on Statistical Shape Modeling

Bio-Imaging Technologies Medical and Scientific Leaders Contribute to the poster presentation for Development Automatic Classification of Vertebral Shape Based on Statistical Shape Modeling at the National Osteoporosis Foundation on April 1-5, 2009 in Washington D.C.

Newtown, PA – April 21, 2009 – Bio-Imaging Technologies announced today that Colin G. Miller, Ph.D. F.I.C.R., Vice President of Business Development and Clinical Affairs, US, is a contributing author of an original article published with the American Medical Association in their Archives of General Psychiatry Publication in the February 2009 issue.

Objective

The purpose of this study was to test a classifier of vertebral fractures based on statistical shape modeling.

Methods

We previously described a method for annotation of vertebral shape using 95 points representing the circumferential vertebral borders on lateral spine x-rays (1), see Figures 1 and 2 below. 
We used this method to manually annotate 2503 vertebrae on lateral thoracic and lumbar spine x-rays from 237 subjects.  Independently, these films were read by an expert radiologist (CH) who classified each vertebra according to the Genant SQ scoring scheme and provided differential diagnosis for non-osteoporotic vertebral deformities.  The image data was as follows: 2390 vertebrae were normal, 967 had non-osteoporotic deformities, and 113 were fractured: 44 fractures were mild (grade 1), 69 were definite (grade 2 or 3).
Using “leave-one-out” methodology we built two-class classifiers to classify all osteoporotic fractures vs non-fractures and specific grade osteoporotic fractures vs non-fractures.

 

Results

The results are presented in Table 1. Balanced error rate classifiers yielded equal sensitivity and specificity of 95.6% overall, 90.1% for mild fractures, and 97.4% for definite fractures with the area under ROC curve of 0.9872, 0.9712 and 0.9975 respectively. The Kappa Score between expert and classifier was 0.64 overall, 0.24 for mild (grade 1), and 0.69 for definite (grade 2 or 3) fractures.

Discussion

These initial results show that a classifier can be constructed to distinguish osteoporotic from normal or non-osteoporotic deformities with an agreement similar to or better than, results published previously for human observers.  Our study showed high sensitivity and specificity to identification of osteoporotic fractures from a set of vertebrae that were either normal or deformed but not by osteoporotic fracture.  Our training set was characterized by relatively few fractures and a high degree of non-osteoporotic deformities.  Building the classifier based on films further enriched for vertebral fracture would likely improve its performance.  Combining this classifier with an efficient annotation tool may provide a useful work flow tool in clinical trials or a decision-making aid in a point-of-care setting.

References

(1) Brett et al. In Proc ASBMR, W227, Hawaii, Sept 2007
Presented at the NATIONAL OSTEOPOROSIS FOUNDATION 8th International Symposium on Osteoporosis: Translating Research into Clinical Practice, Washington, DC, April 1-5, 2009.