A classification and segmentation combined two-stage CNN model for automatic segmentation of brainstem

Abstract

Accurate segmentation of brainstem in MRI images is the basis for treatment of brainstem tumors. It can prevent brainstem from being damaged in neurosurgery. Brainstem segmentation is dominantly based on atlas registration or CNN using patches at present. Nevertheless, the prediction time and the false positive of brainstem segmentation is relatively high. We proposed a classification and segmentation combined two-stage CNN model of brainstem segmentation to improve the prediction accuracy and reduce computation time. Firstly, a classification-CNN model was used to classify MRI images to estimate whether transverse section images exist brainstem. In the view of classified images, a segmentation CNN model to segment brainstem is used to analysis the whole image rather than patches. In addition, considering segmentation based the whole image is a big problem of class unbalance, we settle this problem by changing loss function and giving the label coefficients to get more accurate results. This method provides higher segmentation precision and consume less time for the segmentation task of brainstem than current methods.