To solve problems such as low recognition rate of abnormal cross section and inability to rule out special structural effects, a method based on one-class support vector machine (OCSVM) is proposed to detect coronary lesion. By using coronary cross section resampling and feature selection based on maximum mutual information, the method achieves a relatively high recognition rate. At first, the coronary cross section is resampled based on gradient flux using cubic spline interpolation, and multi-scale feature vector is constructed for every coronary cross section. Then, a maximum mutual information method combined with redundancy removal is adopted to select target features. Finally, selected features are used to train OCSVM model to complete coronary lesion detection. The experiment results in 1128 cross section data show that the maximal recognition rate of the proposed method of health cross section reaches 53.5%, which is much higher than that of support vector machine (SVM) algorithm (learning only from positive and unlabeled data) of 19.6%, with complete recognition of abnormal cross section. Meanwhile, the health cross section recognition rate by 30 features rises from 21.7% to 53.2% owing to the resampling of the cross section.