Effective and efficient multitask learning for brain tumor segmentation
- PDF / 1,685,077 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 69 Downloads / 220 Views
SPECIAL ISSUE PAPER
Effective and efficient multitask learning for brain tumor segmentation Guohua Cheng1 · Jingliang Cheng2 · Mengyan Luo3 · Linyang He3 · Yan Tian4 · Ruili Wang4 Received: 14 November 2019 / Accepted: 17 March 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Recently, brain tumor segmentation has achieved great success, partially because of deep learning-based relation exploration and multiscale analysis. However, the computational complexity hinders the real-time application. In this paper, we propose a revised multitask learning approach in which a lightweight network with only two scales is adopted to segment different kinds of tumor regions. Moreover, we design a hybrid hard sampling method that considers both sample sparsity and effectiveness. Extensive experiments on the BraTS19 segmentation challenge dataset have shown that our proposed method improves the Dice coefficient by a margin of 0.4–1.0 for different kinds of brain tumor regions and obtains results that are competitive with state-of-the-art brain tumor segmentation approaches. Keywords Brain tumor segmentation · Image segmentation · Deep learning · Multitask learning
1 Introduction Brain tumor segmentation, which focuses on separating different tumor regions in multimodel 3D magnetic resonance image (MRI) data, is an important topic for both academia and industry, and it has been an active area of research over the past decade. Effective and efficient brain tumor segmentation can benefit brain disease diagnosis [13], progression assessment [26], and monitoring neurological conditions [25]. Recently, medical image analysis and segmentation have made great progress due to deep learning-based cascade * Guohua Cheng [email protected] 1
Institute of Science and Technology for Brain‑Inspired Intelligence, Ministry of Education‑Key Laboratory of Computational Neuroscience and Brain‑Inspired Intelligence, Fudan University, Shanghai 200433, People’s Republic of China
2
Department of Magnetic Resonance Imaging, The First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, People’s Republic of China
3
Jianpei Technology Co., Ltd, Hangzhou 310018, People’s Republic of China
4
School of Computer and Information Engineering, Zhejiang Gongshang University, Hangzhou 310018, People’s Republic of China
structure networks and multiscale analysis. Cascade methods [11, 23] segment the whole tumor (WT) from the multimodal volumetric data first and then use the WT mask as a prior to constrain the potential region of the tumor core (TC). Again, the TC result is utilized as evidence for enhancing tumor (ET) segmentation. U-Nets [10, 22] extract multiscale feature maps in an encoder–decoder framework with deep residual learning [9]. Though these two kinds of approaches achieve great success in brain tumor segmentation, their computational complexity hinders their real-time application [6, 12, 20, 24]. We illustrate this difficulty in Fig. 1. Here, the Dice coefficient (DC) and floating-point o
Data Loading...
