MRI Brain Tumor Segmentation Combining Saliency and Convolutional Network Features

This paper proposes a brain tumor segmentation method based on visual saliency features on MRI image volumes. The proposed method uses a novel combination of multiple MRI modalities and highlights the potential tumors by applying a healthy template generated from the annotated database slices without tumors. The introduced method proposes a saliency model that includes color and spatial features and as a novel contribution, also incorporates information about the relation of neighboring slices. Based on the saliency map, the outline of the tumor is detected by a region-based active contour method. Moreover, the proposed method is also combined with convolutional neural networks to reduce the networks' eventual overfitting which may result in weaker predictions for unseen cases. By introducing a proof-of-concept method for the fusion of deep learning techniques with saliency-based, handcrafted feature models, the fusion approach has good abstraction skills and yet it is able to handle diverse cases for which the net was less trained. The proposed methods were tested on the BRATS2015 database, and the quantitative results showed that hybrid models (including both trained and handcrafted features) can be promising alternatives to reach higher segmentation performance.