To deal with cancer multi-cancer classification problems, a Multi-class feature selection method based on Support Vector Machine Recursive Feature Elimination (MSVM-RFE) has been proposed. However, it takes the combined weights of all SVM-RFE sub-classifiers into consideration, and ignores the ability of feature selection of each SVM-RFE sub-classifiers. To improve the recognition rate of multi-classification problem, a Modified MSVM-RFE (MMSVM-RFE) was presented. Similar to MSVM-RFE, MMSVM-RFE converted a multi-class problem into multiple binary tasks, then each binary feature elimination problem was solved by an SVM-REF which iteratively removed irrelevant features to obtain a feature subset. All these feature subsets were merged into one final feature subset on which an SVM classifier was trained. The experimental results on three gene datasets show that the proposed method can select a useful feature subset which is efficient in cancer classification. The proposed algorithm can increase the overall recognition rate by about 2%, and significantly enhances the precision of a single category, even to 100%. Compared to random forest, K-Nearest Neighbor (KNN) classifier and PCA dimension reduction, the proposed method can achieve better performance.