| 摘要: |
| :目的 多模态核磁共振成像之间由于包含差异性和互补性信息,在临床实践上往往被综合运用来进行诊断
和脑肿瘤分割。 然而,在真实世界中,可能无法获取到完整模态的脑部核磁成像,而一种或者多种关键模态的缺失
会导致分割性能的下降甚至崩溃。 为了解决这一问题,提出了一种引入基于注意力的解耦再重构模块的分割网
络,希望能够在缺失模态下仍能保持对脑肿瘤的分割精度。 方法 网络使用 U 型架构,并引入基于注意力机制的解
耦再重构模块。 在解耦再重构模块中,首先通过各模态的高维特征融合获得共有基本特征,然后在此基础上借助
差异注意力机制获得各模态的个性特征,最后将基本特征和个性特征经过交叉注意力整合重构后送入解码器中获
取最终分割结果。 同时,还设计了可学习的跳连接门,以使编码器中多种模态的浅层特征经过充分交互后传向解
码器。 结果 在一系列实验中,分割网络在 BraTS2020 脑胶质瘤数据集上表现出了出色的性能,缺失模态下仍能完
成对脑肿瘤的分割,评价指标相较其他先进的方法提高约 2%以上,高达 85. 3%。 结论 因此,基于注意力机制的脑
肿瘤分割网络能够有效地在缺失模态下完成对脑肿瘤的分割,具有重要的实际意义。
关键词:注意力机制;脑肿瘤分割;缺失模态;核磁成像 |
| 关键词: 注意力机制 脑肿瘤分割 缺失模态 核磁成像 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Brain Tumor Segmentation Network Based on Attention Mechanism under Missing Modalities |
|
ZHANG Yuheng LIU Heng
|
|
School of Computer Science and Technology Anhui University of Technology Maanshan 243000 Anhui China
|
| Abstract: |
| Objective Multi-modal magnetic resonance imaging MRI is often used in clinical practice for diagnosis and
brain tumor segmentation due to its containing differential and complementary information. However in real-world
scenarios complete modalities of brain MRI may not always be available and the absence of one or more key modalities
can lead to a decrease or even collapse in segmentation performance. To address this issue a segmentation network
incorporating an attention-based decoupling and reconstruction module is proposed to maintain the segmentation accuracy
for brain tumors even in the presence of missing modalities. Methods The network adopted a U-shaped architecture and
integrated a decoupled and reconstitution module based on attention mechanisms. In this decoupling and reconstitution
module high-dimensional features from each modality were first fused to obtain common basic features. Subsequently by
employing a differential attention mechanism individual features unique to each modality were extracted from the shared
common basic feature. Finally basic features and individual features were integrated through cross-attention and fed into
the decoder to obtain the final segmentation results. Additionally a learnable skip connection gate was designed to allow
the shallow features of multiple modalities in the encoder to interact sufficiently before being passed to the decoder.Results In a series of experiments the segmentation network demonstrated excellent performance on the BraTS2020 brain
glioma dataset successfully segmenting brain tumors even under missing modalities. The evaluation metrics showed an
improvement of over 2% compared with other advanced methods reaching up to 85. 3%. Conclusion Therefore the
attention-based brain tumor segmentation network can effectively segment brain tumors in incomplete modalities which is
of great practical significance. |
| Key words: attention mechanism brain tumor segmentation missing modality magnetic resonance imaging |
