Bidimensional Measurements in Brain Tumors: Assessment of Interobserver Variability

Abstract

Bidimensional tumor measurements indicating a greater than 25% increase in tumor size are generally accepted as indicating tumor progression. We hypothesized that use of digital images and a homogeneous reader population would have lower interobserver variability than in previous studies. Eight board-certified radiologists measured tumor diameters in three planes in two consecutive MRI examinations of 22 patients with contrast-enhancing high-grade brain tumors. Products of tumor measurements were calculated, and determinations were made about tumor progression (> 25% increase in area). A variance components model was run on diameter products and the ratios of consecutive maximal diameter products. The variance components included patient examination effect, reader effect, and residual effect. Complete agreement was found among readers in 10 cases (45%), all indicating stable disease. In the other 12 cases, at least one reader considered progressive disease present. The variance components model showed variance due to readers was small, indicating only modest bias among readers. The residual variance component was large (0.038), indicating that repeated measurements on the same image likely are variable even for the same reader. This variability in measurement implies that repeated measurements by the typical reader have an inherent 14% false-positive rate in the diagnosis of progression of tumors that are stable. Our hypothesis was disproved. We found substantial interreader disagreement and indications that the very nature of the measurement method produces a high rate of false-positive readings of stable tumors. These findings should be considered in interpretation of images with this widely accepted criterion for brain tumor progression.