| 引用本文: | 陈祈剑,王黎,郭顺超,邓泽宇,张健,王丽会.深度多尺度不变特征网络预测胶质瘤1p/19q缺失状态.软件学报,2022,33(12):4559-4573 |
| |
|
| |
|
|
| 本文已被：浏览 1783次 下载 3932次 |
码上扫一扫！ |
|
|
| 深度多尺度不变特征网络预测胶质瘤1p/19q缺失状态 |
|
陈祈剑1,2, 王黎1,2, 郭顺超1,2,3, 邓泽宇1,2, 张健1,2, 王丽会1,2
|
|
1.贵州省智能医学影像分析与精准诊断重点实验室 (贵州大学 计算机科学与技术学院), 贵州 贵阳 550025;2.贵州大学 计算机科学与技术学院, 贵州 贵阳 550025;3.黔南民族师范学院 计算机与信息学院, 贵州 都匀 558000
|
|
| 摘要: |
| 准确预测胶质瘤染色体1p/19q的缺失状态对于制定合适的治疗方案和评估胶质瘤的预后有着重要的意义.虽然已有研究能够基于磁共振图像和机器学习方法实现胶质瘤1p/19q状态的准确预测，但大多数方法需要事先准确勾画肿瘤边界，无法满足计算机辅助诊断的实际需求.因此，提出一种深度多尺度不变特征网络（deep multi-scale invariant features-based network，DMIF-Net）预测1p/19q的缺失状态.首先利用小波散射网络提取多尺度、多方向不变特征，同时基于深度分离转聚合网络提取高级语义特征，然后通过多尺度池化模块对特征进行降维并融合，最后在仅输入肿瘤区域定界框图像的情况下，实现胶质瘤1p/19q状态的准确预测.实验结果表明，在不需要准确勾画肿瘤边界的前提下，DMIF-Net预测胶质瘤1p/19q缺失状态的AUC （area under curve）可达0.92（95% CI=[0.91，0.94]），相比于最优的主流深度学习模型其AUC增加了4.1%，灵敏度和特异性分别增加了4.6%和3.4%，相比于最好的胶质瘤分类前沿模型，其AUC与精度分别增加了4.9%和5.5%.此外，消融实验证明了本文所提出的多尺度不变特征提取网络可以有效地提高模型的预测性能，说明结合深度高级语义特征和多尺度不变特征可以在不勾画肿瘤边界的情况下，显著增加对胶质瘤1p/19q缺失状态的预测能力，进而为低级别胶质瘤的个性化治疗方案制定提供一种辅助手段. |
| 关键词: 胶质瘤 1p/19q 深度学习 小波散射 多尺度不变特征 |
| DOI：10.13328/j.cnki.jos.006499 |
| 分类号:TP181 |
| 基金项目:国家自然科学基金（62161004）；贵州省科学技术基金重点项目（黔科合基础-ZK[2021]重点002）；中法“蔡元培”交流合作项目（2018（No.41400TC））；贵州省科学计划（黔科合基础[2020]1Y255）；贵州省教育厅青年项目（黔教合KY字[2016]321） |
|
| Deep Multi-scale Invariant Features-based Network for Predicting Status of 1p/19q in Glioma |
|
CHEN Qi-Jian1,2, WANG Li1,2, GUO Shun-Chao1,2,3, DENG Ze-Yu1,2, ZHANG Jian1,2, WANG Li-Hui1,2
|
|
1.Key Laboratory of Intelligent Medical Image Analysis and Precise Diagnosis of Guizhou Province (College of Computer Science and Technology, Guizhou University), Guiyang 550025, China;2.College of Computer Science and Technology, Guizhou University, Guiyang 550025, China;3.School of Computer and Information, Qiannan Normal University for Nationalities, Duyun 558000, China
|
| Abstract: |
| Accurately predicting the status of 1p/19q is of great significance for formulating treatment plans and evaluating the prognosis of gliomas. Although there are some works which can predict the status of 1p/19q accurately based on magnetic resonance images and machine learning methods, they require to delineate the tumor contour preliminarily, which cannot satisfy the needs of computer-aided diagnosis. To deal with this issue, this work proposes a novel deep multi-scale invariant features-based network (DMIF-Net) for predicting 1p/19q status in glioma. Firstly, it uses the wavelet-scattering network to extract multi-scale and multi-orientation invariant features, and deep split and aggregation network to extract semantic features. Then, it reduces the feature dimensions using a multi-scale pooling module and fuses these features with concatenation. Finally, with inputting the bounding box of the tumor region it can predict the 1p/19q status accurately. The experimental results illustrate that, without requiring to delineate the tumor region accurately, the AUC predicted by DMIF-Net can reach 0.92 (95%CI=[0.91, 0.94]). Compared with the best deep learning model, the AUC, sensitivity, and specificity increased by 4.1%, 4.6%, and 3.4%, respectively. Compared with the state-of-the-art models on glioma, AUC and accuracy have increased by 4.9% and 5.5%, respectively. Moreover, the ablation experiments demonstrate that the proposed multi-scale invariant feature extraction module can promote effectively the 1p/19q prediction performance, which verify that combining the semantic and multi-scale invariant features can significantly increase the prediction accuracy for 1p/19q status without knowing the boundaries of tumor region, providing therefore an auxiliary means for formulating personalized treatment plan for low-grade glioma. |
| Key words: Glioma 1p/19q deep learning wavelet scattering multi-scale invariant feature |
|
|
|
|
