面向非均衡磁共振影像数据的鲁棒病脑检测方法研究

It is a key research topic for pathological brain detection how to remove the effect from electronic interference and noises. Low-rank decomposition theory (LRDT) suggests that practical observations can be treated as a combination of a low-rank component and a sparse noise component. Accordingly, scholars can reconstruct original clean data from samples corrupted by noises. Nevertheless, the assumption in LRDT takes a hypothesis that the noises follow either Gaussian distribution or Laplacian distribution, which does not hold on the scene of magnetic resonance imaging (MRI) scanning. Another problem is the unbalanced data obtained by MRI. To solve above two problems, this grant proposed a noise adaptive robust feature extraction and robust classification methods for unbalanced MRI data based on LRDT, in order to detect pathological brains. Our method can adaptively filter the corruption of the noises, no matter what types of electronic interference and noises are. This grant focused on robust feature extraction and robust classification methods oriented for unbalanced brain MRI data, which can extend the theoretical system of pattern recognition and artificial intelligence, and can enrich the application scope of computer-aided diagnosis. This grant can give better technical support for robust pathological brain detection in MRI scanning.

如何在核磁共振扫描环境下，过滤电子干扰与噪声的影响，是病脑检测技术面临的难点。低秩分解理论表征：实际观测量可归纳为低秩分量与稀疏噪声分量的结合。因此，研究者可从噪声污染的样本中恢复原始的干净数据。然而，低秩分解理论存在如下假设：即噪声满足高斯或拉普拉斯分布。该假设在磁共振扫描场景中不成立、且脑影像数据往往是非均衡的。为了解决上述两个问题，本项目提出基于低秩分解理论的面向非均衡磁共振影像数据的鲁棒特征提取方法和鲁棒分类方法，用于鲁棒病脑检测。我们的方法可在不同电子干扰与噪声条件下，自适应地消除掉脑磁共振影像数据中的噪声污染。本项目重点研究面向非均衡脑磁共振数据的鲁棒特征抽取与分类方法，进一步扩充了模式识别与人工智能的理论体系，丰富了计算机自动检测的应用范围。本项目可为磁共振扫描环境下的鲁棒病脑检测提供更好的技术支撑。

疾病的早期自动诊断对于病人，医院，社会有重要的经济与社会意义。我国脑相关疾病死亡占总死亡人数的25%以上，我国脑疾病负担达到每年上万亿元。本项目主要研究面向非均衡磁共振影像数据的鲁棒特征提取方法和鲁棒分类方法用于鲁棒病脑检测。项目对脑磁共振图像的电子噪声建模，设计相关的模型参数学习算法以及参数优化算法，给出基于低秩分解的鲁棒分类模型及优化求解算法，使得分类器对环境噪声具备更好的鲁棒性，解决非均衡样本和电子噪声条件下的鲁棒特征提取和分类问题，提高病脑检测的准确率和有效性。额外的，项目引入最新的深度学习算法，包括卷积神经网络，迁移学习，自动编码机等，使算法在病脑分类问题上检测性能进一步提升。我们的算法在多类脑疾病(老年痴呆，多并发硬化，精神分裂，自闭症，听力损失，酒精中毒脑等)上获得了较好的临床验证。