Wang Huaizhi, Liu Haichao, Du Fang, Wang Di, Huo Xianhao, Tian Jihui, Song Lijuan
School of Information Engineering, Ningxia University, Yinchuan, Ningxia, China.
Ningxia Key Laboratory of Artificial Intelligence and Information Security for Channeling Computing Resources from the East to the West, School of Information Engineering, Ningxia University, Yinchuan, Ningxia, China.
Front Oncol. 2025 May 19;15:1574861. doi: 10.3389/fonc.2025.1574861. eCollection 2025.
Glioma stands as one of the most lethal brain tumors in humans, and its accurate diagnosis is critical for patient treatment and prognosis. Magnetic Resonance Imaging (MRI) has been widely utilized for glioma diagnosis and research due to its non-invasive nature and clinical accessibility. According to the 2021 World Health Organization Central Nervous System Tumor Classification guidelines, glioma subtypes can be determined through molecular status information of Isocitrate Dehydrogenase (IDH), Chromosome 1p/19q codeletion (1p/19q), and Alpha Thalassemia/Mental Retardation Syndrome X-linked (ATRX) genes.
In this study, we propose a dual-path parallel fusion network (MDPNet) designed to comprehensively extract heterogeneous features across different MRI modalities while simultaneously predicting the molecular status of IDH, 1p/19q, and ATRX. To mitigate the impact of data imbalance, we developed a cross-gene feature-sharing classifier and implemented an adaptive weighted loss function, substantially enhancing the model's predictive performance.
In this study, each gene classification task was formulated as a binary classification problem. Experiments conducted on public datasets demonstrate that our method outperforms existing approaches in accuracy, Area Under the Curve (AUC), sensitivity, and specificity. The achieved classification accuracies for IDH, ATRX, and 1p/19q reach 86.7%, 92.0%, and 89.3%, respectively. The source code of this study can be viewed at https://github.com/whz847/MDPNet.
The proposed framework exhibits significant advantages in integrating heterogeneous features from multi-modal MRI data. Experimental results from internal datasets further validate the model's superior generalizability and clinical utility in assisting glioma diagnosis, highlighting its potential for real-world clinical applications.
胶质瘤是人类最致命的脑肿瘤之一，其准确诊断对患者的治疗和预后至关重要。磁共振成像（MRI）因其非侵入性和临床可及性，已被广泛用于胶质瘤的诊断和研究。根据2021年世界卫生组织中枢神经系统肿瘤分类指南，胶质瘤亚型可通过异柠檬酸脱氢酶（IDH）、染色体1p/19q共缺失（1p/19q）和X连锁α地中海贫血/智力发育迟缓综合征（ATRX）基因的分子状态信息来确定。
在本研究中，我们提出了一种双路径并行融合网络（MDPNet），旨在全面提取不同MRI模态的异构特征，同时预测IDH、1p/19q和ATRX的分子状态。为了减轻数据不平衡的影响，我们开发了一种跨基因特征共享分类器，并实现了自适应加权损失函数，显著提高了模型的预测性能。
在本研究中，每个基因分类任务都被制定为一个二分类问题。在公共数据集上进行的实验表明，我们的方法在准确率、曲线下面积（AUC）、敏感性和特异性方面优于现有方法。IDH、ATRX和1p/19q的分类准确率分别达到86.7%、92.0%和89.3%。本研究的源代码可在https://github.com/whz847/MDPNet查看。
所提出的框架在整合多模态MRI数据的异构特征方面具有显著优势。内部数据集的实验结果进一步验证了该模型在辅助胶质瘤诊断方面的卓越通用性和临床实用性，突出了其在实际临床应用中的潜力。
