Thung Kim-Han, Yap Pew-Thian, Shen Dinggang
Department of Radiology and BRIC, University of North Carolina, Chapel Hill, USA.
Deep Learn Med Image Anal Multimodal Learn Clin Decis Support (2017). 2017 Sep;10553:160-168. doi: 10.1007/978-3-319-67558-9_19. Epub 2017 Sep 9.
Utilization of biomedical data from multiple modalities improves the diagnostic accuracy of neurodegenerative diseases. However, multi-modality data are often incomplete because not all data can be collected for every individual. When using such incomplete data for diagnosis, current approaches for addressing the problem of missing data, such as imputation, matrix completion and multi-task learning, implicitly assume linear data-to-label relationship, therefore limiting their performances. We thus propose multi-task deep learning for incomplete data, where prediction tasks that are associated with different modality combinations are learnt jointly to improve the performance of each task. Specifically, we devise a multi-input multi-output deep learning framework, and train our deep network subnet-wise, partially updating its weights based on the availability of modality data. The experimental results using the ADNI dataset show that our method outperforms the state-of-the-art methods.
利用来自多种模态的生物医学数据可提高神经退行性疾病的诊断准确性。然而，多模态数据往往是不完整的，因为并非每个个体都能收集到所有数据。在使用此类不完整数据进行诊断时，当前解决数据缺失问题的方法，如插补、矩阵补全和多任务学习，隐含地假设了数据与标签之间的线性关系，因此限制了它们的性能。因此，我们提出了针对不完整数据的多任务深度学习方法，其中与不同模态组合相关的预测任务被联合学习，以提高每个任务的性能。具体而言，我们设计了一个多输入多输出深度学习框架，并按子网训练我们的深度网络，根据模态数据的可用性部分更新其权重。使用ADNI数据集的实验结果表明，我们的方法优于现有方法。
