Multimodal brain tumor image segmentation using WRN-PPNet. (July 2019)

• Record Type: Journal Article
• Authors: Wang, Yu; Li, Changsheng; Zhu, Ting; Zhang, Jingyang
• Is Part Of: Computerized medical imaging and graphics. Volume 75(2019)
• Page Start: 56
• Page End: 65
• Publication Date: 2019-07
• Subjects: Multi-modality magnetic resonance imaging -- Gliomas -- Wide residual net -- Pyramid pool net -- End to end -- Automatic segmentation; Diagnostic imaging -- Periodicals; Imaging systems in medicine -- Periodicals; Diagnosis, Radioscopic -- Data processing -- Periodicals; Diagnostic Imaging -- Periodicals; Imagerie pour le diagnostic -- Périodiques; Diagnostic imaging; Periodicals; Electronic journals; Electronic journals; 616.0754
• DOI: 10.1016/j.compmedimag.2019.04.001
• Languages: English
• ISSNs: 0895-6111

Abstract

Highlights: WRN-PPNet model is proposed firstly to extract features from multimodal MRI brain tumor images. Different level global prior representation of slices is obtained through pyramid pooling module. Global prior representation is combined with original inputs for brain tumor segmentation. The process to segment the brain tumor is fully automatically, end-to-end, robust and fast. Abstract: Tumor segmentation is of great importance for diagnosis and prognosis of brain cancer in medical field. Because of the noise, inhomogeneous gray, diversity of tissue, bias among modalities, and the fuzzy boundaries between tumor and adjacent tissues in the magnetic resonance imaging (MRI), tumor segmentation is a very difficult task. At present, many of the existing brain tumor segmentation methods are semi-automatic which are troublesome and inconvenient because interventions of raters or specialists are required. In this paper an automatic method, named wide residual network & pyramid pool network (WRN-PPNet), which can automatically segment glioma end to end is put forward. The main idea is described below. Firstly, WRN is used to extract features of multimodal brain tumor slices which are proved to have strong expressive ability. Secondly, the global prior representation with different level obtained by PPNet is stacked on the features from WRN. Finally, the scale recovery module in which the original inputs are fed into the network again is utilized to produce the pixel-level predictions which have the same size with the original inputs. Dice coefficients, sensitivity coefficient and predictive positivity value (PPV) coefficient are used to evaluate the performance of WRN-PPNet quantitatively. The experimental results show that the proposed method has a superior performance compared with the other state-of-the-art methods, and the average Dice, sensitivity and PPV on the randomly selected test data can reach 0.91, 0.94 and 0.89 respectively.