Quantitation of T2 lesion load in multiple sclerosis with magnetic resonance imaging: a pilot study of a probabilistic neural network approach.

Abstract

To quantitate multiple sclerosis (MS) lesions in the brain by using computerized techniques. MS lesions from five patients were quantitated with magnetic resonance (MR) imaging by using three approaches: a probabilistic neural network (PNN) approach, a semiautomated method that uses a bifeature space approach with operator intervention at each section, and the "gold standard" of manual outlining of lesions. Each patient underwent two MR studies in 1 day. The PNN approach allows reasonable quantitation of large data sets with minimal operator input. The mean intraobserver error for the PNN approach was competitive with the more time-consuming bifeature space approach (5.2% vs 4.4%, respectively). On average, both computer assisted methods performed better than the manual method (mean intraobserver error, 10.1%). The agreement between the two computerized quantitation approaches was good. The number of interactive steps was substantially reduced with the PNN technique, leading to minimal operator intervention time.