One reason for the success of deep learning in natural image data is the availability of large-scale labeled data. However, labeled medical image data often is limited, as annotating medical image data requires extensive human efforts and expertise. Consequently, any variation and uncertainty in medical image data would affect the training process, and the capacity of deep learning usually cannot be fully explored. In this thesis, we propose to improve deep learning performance by addressing variations and uncertainties in medical image data.

First, we study the inter-observer problem in retinal vessel segmentation, where human observers can generate different pixel-wise annotations given the same medical image. To address the problem, we design a vessel thickness similarity me...[ Read more ]

One reason for the success of deep learning in natural image data is the availability of large-scale labeled data. However, labeled medical image data often is limited, as annotating medical image data requires extensive human efforts and expertise. Consequently, any variation and uncertainty in medical image data would affect the training process, and the capacity of deep learning usually cannot be fully explored. In this thesis, we propose to improve deep learning performance by addressing variations and uncertainties in medical image data.

First, we study the inter-observer problem in retinal vessel segmentation, where human observers can generate different pixel-wise annotations given the same medical image. To address the problem, we design a vessel thickness similarity measure and construct a segment-level loss function accordingly. Then, we integrate the loss function with a two-branch deep learning framework. Experiments on publicly available datasets demonstrate the effectiveness of our approach.

Second, we investigate the boundary uncertainty problem in gland instance segmentation. Due to limited image resolution, annotated boundaries are not always correct, which makes it challenging to preserve shape information. To relax the constraint for boundary detection, we propose a shape-aware adversarial learning framework to enable one single deep learning model for accurate gland instance segmentation. Our evaluations confirm that our method can obtain better performance compared with other state-of-the-art methods.

Lastly, we discuss the cross-client variation problem, where image data from different sources can vary significantly. It will be the bottleneck when applying federated learning to train deep learning models from multi-source decentralized medical image data. We, for the first time, propose a variation-aware federated learning framework to address the problem. Experimental results on the classification of clinically significant prostate cancer from multi-source decentralized ADC image data show that our framework outperforms the current federated learning framework, especially when dealing with small datasets.