Neural networks have revolutionized the field of computer vision since they provide solutions to a number of previously unsolved problems and achieve promising performance both in terms of accuracy and computational efficiency. It has increasingly become recognized as providing high performance for applications as diverse as image classification, object detection, image segmentation, compressive sensing, inpainting, action recognition, and super resolution. In this dissertation, we design neural network models for various image segmentation tasks, including chronic stroke lesion segmentation and semi-supervised video segmentation. In video segmentation, the proposal of one shot learning contributes greatly to object segmentation. The state-of-the-art segmentation methods demonstrate the great benefit of combining offline training and online \textit{one shot finetuning}, which updates the weights of the whole model to guide it towards a target object. Unfortunately, this technique has very limited practical use both in terms of computational cost and strict memory demand. In our work, we present a network modulator which is an add-on component of a general segmentation network specifically designed to eliminate the shortcoming of traditional two-stage video segmentation scheme. Network modulator allows for quick adaption of a segmentation net towards a target object via a surprisingly simple layer-wise transformation. The layer-wise transformation consists of scaling and shifting operations on each feature map of the model. The network modulators are generalized to any fully convolutional networks (FCN) based applications; this work focuses on semi-supervised video segmentation problems. Experiments show that the network modulation leads to good results and efficient evaluations. Chronic stroke lesion segmentation on magnetic resonance imaging (MRI) scans plays a critical role in helping physicians to determine stroke patient prognosis. Previous segmentation studies have used statistical machine learning and neural networks to address the task, but the important feature - brain symmetry property - has been mostly ignored in the past studies. In our work, we show the benefit of a neural network that combines physiologically based information, that is, the brain symmetry. One of the main challenges that limits the use of the brain symmetry is its complexity: \textit{the brain is not perfectly symmetric at the pixel-level}. We propose a convolutional neural network (CNN) segmentation network - a 3D Cross-hemisphere Neighborhood Difference ConvNet - which is specifically designed for brain symmetry. The main novelty of this network lies in a 3D cross-hemisphere neighborhood difference layer which introduces robustness to position and scale in brain symmetry. Such robustness is important in helping the CNN distinguish between minute hemispheric differences and the asymmetry caused by a lesion. Experiment comparison with the state-of-the-art methods demonstrates the effectiveness of the proposed model. In this dissertation we demonstrate that the simple layer-wise modification is surprisingly effective and critical in adapting the network into the target domain. We hope our work will provide insights for further investigations on the nature of neural network. In addition, most of the proposed computational methods in medical image prediction were validated on relatively small private datasets, making objective comparisons between methods highly challenging. To tackle this challenge, we built a benchmark multi-class chronic stroke dataset for white matter hyperintensity and infarct lesion segmentation. We hope this densely annotated, pixel-accurate, and per-slice ground-truth segmentation dataset will provide a benchmark for training and testing lesion segmentation algorithms and will be a standardized dataset for comparing the performance of different computational methods.

Last modified

• 09/10/2019

Creator

• Yan-Ran Wang

DOI

• https://doi.org/10.21985/n2-q7e1-m530

Subject

• Electrical Engineering
• Computer Science

Keyword

• video segmentation
• neural network
• medical image analysis
• chronic stroke lesion segmentation

Date created

• 2019

Rights statement

• In Copyright