Baker's Cyst Classification Using Random Forests

In this paper, a classification procedure for Baker's cysts was proposed. The procedure contained two subprocedures: the image preprocessing (dual thresholding, labeling, feature extraction) and the classification (Random Forests, cross validation). In total, five features were required to classify the cysts. These geometric features represented the location, the area and the convexity of the cyst. The procedure was proven effective on a set 436 varied MRI images. The set contained 68 images with cysts ready for aspiration and was oversampled with the SMOTE approach. The proposed method operates on 2D MRI images. This reduces the time of diagnosis and, with the ever increasing demand for MRI scanners, is justified economically. The method can be employed in systems for autonomous and semi-autonomous Baker's cyst aspiration or as a standalone package for MRI images annotation. Furthermore, it can be also extended to other fluid-based medical conditions in the knee.