Traditional methods usually use corners and extreme points as feature points

and ignore the changes of texture so that the performances of the medical image classification are affected. A new feature point detection and description method is provided for medical image classification task using the Bag-of-Keypoints model. First

adaptive K-means is used to cluster images on the pixel-level

and the points where the clustering distribution in its univalue segment assimilating nucleus (USAN) changes rapidly are selected to be the features points. Second

the descriptor is defined in a polar coordinate system and a virtual dictionary is constructed in order to describe the image by virtual word histogram. Last

histogram intersection is used to measure the similarity between images

and K-nearest neighbor (KNN) is extended by it to finish the classification. Image retrieval in medical applications (IRMA)medical image classification code is strictly followed when experimental data is selected. The results show that the average value increased 4.5% than traditional classification; our method is more stable and robust for different classes of image

and meets the needs of clinical application better.