Cerebral Metastases Segmentation using Transfer Gliomas Learning and GrabCut

Abstract

Segmentation of medical images is an important area of research that can be used in prognosis prediction and patient treatment. Due to the high variability of data, the task to develop an accurate segmentation method remains challenging. In this paper we address the problem of cerebral metastases segmentation and focus our analysis on the BrainMetShare dataset. In order to enhance the metastases segmentation performance we propose a two stage method. In the first stage we employ a transfer learning procedure where we train an Unet model on the similar task of low and high grade gliomas segmentation provided by the BraTS dataset and then fine-tune the model for solving our problem of cerebral metastases segmentation. In the second stage we use GrabCut to refine the metastases segmentation masks obtained from the first stage. In the experimental evaluation we show that our two stage method based on transfer learning and GrabCut progressively outperforms the baseline model trained only on cerebral metastases data from BrainMetShare.