南方医科大学学报

南方医科大学学报 ›› 2023, Vol. 43 ›› Issue (8): 1379-1387.doi: 10.12122/j.issn.1673-4254.2023.08.15

• • 上一篇    下一篇

磁共振酰胺质子转移模态的胶质瘤IDH基因分型识别：基于深度学习的Dual-Aware框架

楚智钦，屈耀铭，钟 涛，梁淑君，温志波，张 煜   

  1. 南方医科大学生物医学工程学院，广东省医学图像处理重点实验室，广东 广州 510515；南方医科大学珠江医院放射科，广东 广州 510282
  • 出版日期:2023-08-20 发布日期:2023-09-11

A Dual-Aware deep learning framework for identification of glioma isocitrate dehydrogenase genotype using magnetic resonance amide proton transfer modalities

CHU Zhiqin, QU Yaoming, ZHONG Tao, LIANG Shujun, WEN Zhibo, ZHANG Yu   

  1. School of Biomedical Engineering, Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou 510515, China; Department of Radiology, Zhujiang Hospital, Southern Medical University, Guangzhou 510282, China
  • Online:2023-08-20 Published:2023-09-11

RichHTML

10

PDF

151

摘要： 目的 基于磁共振酰胺质子转移（APT）模态数据，提出Dual-Aware深度学习框架实现胶质瘤异柠檬酸脱氢酶（IDH）基因分型，实现无创性的辅助诊断。方法 共收集118例多模态磁共振脑影像数据，其中68例为野生型，50例为突变型，所有数据完成了脑部胶质瘤ROI区域勾画。APT模态成像无需造影剂，它对肿瘤的信号强度跟肿瘤恶性程度呈正相关，且对IDH野生型的信号强度高于对IDH突变型的信号强度。针对APT模态肿瘤成像及衍生区域形态多变，且相对其他模态更缺乏明显的边缘轮廓特点，构建Dual-Aware框架，引入Multi-scale Aware模块挖掘多尺度特征，Edge Aware模块挖掘边缘特征，以实现基因分型的自动识别。结果 两类Aware机制的引入能够有效提升模型对胶质瘤IDH基因分型的识别率。对各模态数据在准确率和AUC方面都取得了一定的提升，并在APT模态上取得了最佳性能，预测精度达到83.1%，AUC达到0.822，从而验证了APT模态对于脑胶质瘤IDH基因分型识别的优势和有效性。结论 本文采用的深度学习算法模型基于APT模态的图像特点构建，验证了APT模态在胶质瘤IDH基因分型识别任务中的有效性，有望在该任务利用APT模态取代目前常用的T1CE模态，避免造影剂注射，以实现无创性的IDH基因分型。

关键词: 深度学习；Dual-Aware；酰胺质子转移；胶质瘤；异柠檬酸脱氢酶

Abstract: Objective To propose a Dual-Aware deep learning framework for genotyping of isocitrate dehydrogenase (IDH) in gliomas based on magnetic resonance amide proton transfer (APT) modality data as a means to assist non-invasive diagnosis of gliomas. Methods We collected multimodal magnetic resonance imaging (MRI) imaging data of the brain from 118 cases of gliomas, including 68 wild-type and 50 mutant type cases. The delineation of the ROI of brain glioma was completed in all the cases. APT modality imaging does not require contrast agents, and its signal intensity on tumors is positively correlated with tumor malignancy, and the signal intensity on wild- type IDH is higher than that on mutant IDH. For APT modalities, tumor imaging and derived areas are morphologically variable and lack prominent edge contour characteristics compared with other modalities. Based on these characteristics, we propose the Dual-Aware framework, which introduces the Multi-Aware framework to mine multi-scale features, and the Edge Aware module mines the edge features for automatic genotype identification. Results The introduction of two types of Aware mechanisms effectively improved the identification rate of the model for glioma IDH genotyping. The accuracy and AUC for each modality data were enhanced, and the best performance was achieved on the APT modality with a prediction accuracy of 83.1% and an AUC of 0.822, suggesting its advantages and effectiveness for identifying glioma IDH genotypes. Conclusion The proposed deep learning algorithm model constructed based on the image characteristics of the APT modality is effective for glioma IDH genotyping and identification task and may potentially replace the commonly used T1CE modality to avoid contrast agent injection and achieve non- invasive IDH genotyping.

Key words: deep learning; Dual-Aware; amide proton transfer; glioma; isocitrate dehydrogenase