团队研究生李帅发表论文Multi-task parallel with feature sharing integrated 3D U-Nets for Glioma Segmentation

华南理工大学

2024-03-13 22:07:49

2024-03-13

该论文发表于期刊Biomedical Signal Processing and Control，根据脑胶质多序列MR影像的特性，设计了一种并行多通道与特征融合的深度学习网络，实验结果证明该网络在同时分割肿瘤坏死区、增强区和水肿区三个重要区域，均具有优异的性能。

Glioma segmentation is a crucial task for accurate quantitative analysis, precise diagnosis and effective

treatment. However, the challenge remains in simultaneously locating multiple sub-regions within the tumor

from multi-sequence MRI images, as they exhibit heterogeneous appearance, shape, diverse sizes, locations,

and different intensities. The mainstream approach currently used is to use a single network to simultaneously

perform segmentation of gliomas and their internal regions. However, this approach overlooks the differences

between segmentation tasks, leading to the shared use of all intermediate features for all tasks. The effectiveness

of certain features for a specific task relies on the learning of network parameters, lacking structural division.

As a result, the overall performance of joint image segmentation is compromised. To address this issue, a

multi-task parallel with feature sharing integrated 3D U-Nets is proposed, which employs three sub-networks

to identify three sub-regions of the tumor. Each sub-network is dedicated to identifying a specific sub-region of

the tumor, thus dividing various types of features required by different tasks into different sub-networks, and

avoiding mutual interference between tasks. The sub-networks incorporate special feature sharing pathways in

the encoder, allowing them to capture the spatial inclusion relationship and intensity continuity among the subregions.

Furthermore, a compound loss function is defined, as a weighted sum of a Dice loss and a cross entropy

loss to achieve a tradeoff between alleviating class imbalance and promoting training smoothness. Ablation

studies have shown the effectiveness of parallel structure and feature sharing pathways in the proposed method.

Experimental results and comparisons on quantitative and qualitative evaluation demonstrate the proposed

method is superior to classical and state-of-the-art methods. On the BraTS 2021 dataset, the proposed method

achieves the combined optimization of segmentation performance, such as the average Dice, PPV, Sensitivity,

and HD95 over multiple regions achieves the best of 0.908, 0.923, 0.931, 3.312.

登录用户可以查看和发表评论，请前往登录或注册。