On 2 separate occasions at least 1 week apart, 6 independent observers studied errors in identification of the transepicondylar axis (TEA) using computed tomography in 10 cadaveric knees. The reference surgical TEA was established by dissection of specimens. We found an average error of 2.6 degrees (external rotation) when the TEA was identified by computed tomography. The error was significantly smaller when the TEA was identified by using a conventional axial image than when a 3-dimensional reconstruction image was used (P < .001, paired t test). No significant intraobserver error (P = .814, Wilcoxon rank test) was found when the measurement was performed using conventional 2-dimensional computed tomography images. However, significant interobserver difference was evident (P < .001, Friedman test).