改进MobileViT模型的脑胶质瘤良恶分类方法

A Glioma Malignancy Grading Method Based on an Improved MobileViT Model

摘要: 脑胶质瘤良恶性分级对术前个体化诊疗方案的制定具有重要临床价值。磁共振成像(MRI)是胶质瘤无创诊断的首选工具，但人工阅片受主观因素影响，诊断一致性难以保证。针对现有深度学习方法存在的感受野受限、通道特征自适应建模不足及小样本泛化能力弱等问题，本文提出一种基于改进MobileViT的脑胶质瘤良恶性分类方法。在MobileViT骨干网络基础上，引入空洞卷积金字塔模块(Efficient Spatial Pyramid module, ESP模块)构建多膨胀率并行分支与层级特征融合结构，增强对病灶多尺度上下文特征的捕获能力；引入通道注意力模块(Channel Attention module, CA模块)自适应生成通道注意力权重，强化判别性通道特征响应；并结合学习率线性衰减、标签平滑及医学图像迁移学习三种策略优化训练过程。在BraTS2019数据集上的病例级五折交叉验证结果表明，本文方法ACC达98.87%，SEN、SPE、PPV、NPV分别为98.21%、99.53%、99.52%及98.23%，模型参数量仅7.6 M，综合性能优于MobileNetV2、ResNet101、EfficientNetB3等经典网络及现有同类方法。

Abstract: Accurate preoperative malignancy grading of glioma is essential for individualized treatment planning. Although magnetic resonance imaging (MRI) is the preferred modality for non-invasive glioma diagnosis, manual interpretation is susceptible to subjective bias and inconsistency. To address the limitations of existing deep learning methods, including constrained receptive fields, insufficient channel feature modeling, and poor small-sample generalization, this paper proposes an improved MobileViT-based glioma grading method. An Efficient Spatial Pyramid (ESP) module is introduced for multi-scale contextual feature extraction, and a Channel Attention (CA) module is incorporated to reinforce discriminative channel responses. Three training strategies—linear learning rate decay, label smoothing, and medical image-based transfer learning—are further applied. Five-fold cross-validation on BraTS2019 yields an accuracy (ACC) of 98.87%, with sensitivity (SEN), specificity (SPE), positive predictive value (PPV), and negative predictive value (NPV) of 98.21%, 99.53%, 99.52%, and 98.23%, respectively. With only 7.6 M parameters, the proposed method outperforms MobileNetV2, ResNet101, and EfficientNetB3, achieving a favorable balance between lightweight design and classification performance.