There has been increasing interest in quantitatively analyzing diffusion anisotropy of ischemic lesions from diffusion tensor magnetic resonance imaging (DT-MRI). In this study, we develop and evaluate a novel method to automatically segment cerebral ischemic lesions from DT-MRI images. The method is a combination of image preprocessing, measures of diffusion anisotropy, multi-scale statistical classification (MSSC), and partial volume reclassification (PVRC). First, non-linear filtering is applied to DT-MRI images to reduce image noise. Then, measures of diffusion anisotropy are calculated to acquire the diffusion properties of different brain tissues. Finally, ischemic lesions are accurately segmented using robust MSSC-PVRC, taking into account spatial information, intensity gradient, radio frequency (RF) inhomogeity and measures of diffusion anisotropy of DT-MRI images. After MSSC, PVRC is applied to overcome partial volume effect (PVE). Results show that the method got a satisfied segmentation of ischemic lesions, successfully overcoming the problem of intensity overlapping and reducing PVE, and that the method is robust to varying starting parameters. The proposed automatic technique is promising not only to detect the site and size of ischémic lesions in stroke patients but also to quantitatively analyze diffusion anisotropy of lesions for further clinical diagnoses and therapy.