Validating Nonlinear Registration to Improve Subtraction Images for Lesion Detection and Quantification in Multiple...

To propose and validate nonlinear registration techniques for generating subtraction images because of their ability to reduce artifacts and improve lesion detection and lesion volume quantification.

Postcontrast T₁ -weighted spin echo and T₂ -weighted dual echo images were acquired for 20 patients with relapsing-remitting multiple sclerosis (RRMS) on a monthly basis for a year (14 women, average age 33.6 ± 6.9). The T₂ -weighted images from the first scan were used as a baseline for each patient. The images from the last scan were registered to the baseline image. Four different registration algorithms used for evaluation included; linear, halfway linear, nonlinear, and nonlinear halfway. Subtraction images were generated after brain extraction, intensity normalization, and Gaussian blurring. Lesion activity changes along with identified artifacts were scored on all four techniques by two independent observers. Additionally, quantitative analysis of the algorithms was performed by estimating the volume changes of simulated lesions and real lesions. For real lesion volume change analysis, five subjects were selected randomly. Subtraction images were generated between all the 11 time points and the baseline image using linear and nonlinear registration for the five subjects.

Lesion activity detection resulted in similar performance among the four registration techniques. Lesion volume measurements on subtraction images using nonlinear registration were closer to lesion volume on T₂ -weighted images. A statistically significant difference was observed among the four registration techniques while evaluating yin-yang artifacts. Pairwise comparisons showed that nonlinear registration results in the least amount of yin-yang artifacts, which are significantly different.

Nonlinear registration for generation of subtraction images has been demonstrated to be a promising new technique as it shows improvement in lesion activity change detection. This approach decreases the number of artifacts in subtraction images. With improved lesion volume estimates and reduced artifacts, nonlinear registration may lead to discarding less subject data and an improvement in the statistical power of subtraction imaging studies.