The deep learning–based computer-aided diagnosis (CADx) approaches of dementia often require a lot of manual intervention. Although deep learning has a good effect on feature extraction, the current deep learning methods usually need to set a large number of parameters manually, which is time consuming. Hierarchical extreme learning machine (H-ELM) needs only less manual intervention and can extract features by a multi-layer feature representation framework, which is much faster than the traditional deep learning methods. A CADx framework based on H-ELM, named DCADx, is proposed. As common spatial pattern (CSP) and brain functional network (BFN) have been proven to have better de-redundancy effects on brain data, the DCADx contains two different data redundancy reduction methods: (1) CSP-based DCADx (i.e., DCADx-CSP model) and (2) BFN-based DCADx (i.e., DCADx-BFN model). The experimental evaluation proved the effectiveness of the proposed algorithms. The DCADx-CSP model obtained 83.2% on Alzheimer’s disease and 82.5% on Parkinson’s disease. The DCADx-BFN obtained 89.3% on Alzheimer’s disease and 88.7% on Parkinson’s disease. DCADx can make full use of the feature expression ability of H-ELM to achieve better performance. CSP and BFN can reduce the redundancy to enhance the diagnostic accuracy further.

基于深度学习的痴呆症计算机辅助诊断（CADx）方法通常需要大量的人工干预。尽管深度学习对特征提取有很好的效果，但是当前的深度学习方法通​​常需要手动设置大量参数，这很耗时。分层极限学习机（H-ELM）仅需要较少的人工干预，并且可以通过多层特征表示框架提取特征，这比传统的深度学习方法快得多。提出了一种基于H-ELM的CADx框架DCADx。由于已证明常见的空间模式（CSP）和大脑功能网络（BFN）对大脑数据具有更好的去冗余效果，因此DCADx包含两种不同的数据冗余减少方法：（1）基于CSP的DCADx（即，DCADx-CSP模型）和（2）基于BFN的DCADx（即DCADx-BFN模型）。实验评估证明了所提算法的有效性。DCADx-CSP模型获得83.2％的阿尔茨海默氏病和82.5 ％的帕金森氏病。DCADx-BFN的阿尔茨海默氏病占89.3 ％，帕金森氏病占88.7 ％。DCADx可以充分利用H-ELM的特征表达能力来获得更好的性能。CSP和BFN可以减少冗余以进一步提高诊断准确性。