Although well-established deep learning models such as convolutional neural networks have achieved wide success in the natural image domain, there remain lots of challenges in developing reliable deep learning-based diagnostic models due to the imperfect nature of medical image data. Publicly available large-scale medical datasets with annotations are very limited. The various protocols and inherent noises in the imaging systems also make the representation learning more challenging. In this thesis, we propose three works to address the challenges in deep learning-based disease diagnostic models as follows.

Firstly, we investigate the histopathologic detection problem on high-resolution histology images, where directly training a deep CNN in image-level is computationally infeasible. Do...[ Read more ]

Although well-established deep learning models such as convolutional neural networks have achieved wide success in the natural image domain, there remain lots of challenges in developing reliable deep learning-based diagnostic models due to the imperfect nature of medical image data. Publicly available large-scale medical datasets with annotations are very limited. The various protocols and inherent noises in the imaging systems also make the representation learning more challenging. In this thesis, we propose three works to address the challenges in deep learning-based disease diagnostic models as follows.

Firstly, we investigate the histopathologic detection problem on high-resolution histology images, where directly training a deep CNN in image-level is computationally infeasible. Down-sampling is not an optimal strategy either, as the local regions contain discriminative details for identifying carcinoma. To address this challenge, we propose a deep spatial fusion network that can preserve the local details while exploiting the global features in representation learning, which significantly improves the diagnostic accuracy for breast cancer.

Secondly, we study the problem of cancerous region localization using weakly supervised learning. We propose a new attention-based neural network and a localization method that learns to localize the evidence supporting the diagnostic decision without requiring object-level annotations, which eases the intensive annotation labor and make the diagnostic model more interpretable. Comprehensive experiments are conducted to evaluate the effectiveness of our method.

Lastly, we address the challenge of population-based disease prediction on multimodal data. We propose an edge-variational graph convolutional network that, on the one hand, adaptively constructs a population graph by estimating the association between subjects, on the other hand, performs semi-supervised disease prediction with uncertainty estimation using graph learning. Extensive experiments show that our approach can complementarily combine imaging and non-imaging data to improve the disease prediction performance on Autism Spectrum Disorder, Alzheimer's Disease, and ophthalmic diseases.