THESIS
2020
xvii, 99 pages : illustrations ; 30 cm
Abstract
Deep learning has achieved great success in different tasks for natural image data in recent
years. However, the segmentation of medical image data appears to be challenging because
of the limited manually labeled data from the medical experts, the pathological changes
and the morphological variation of the target objects, and the random noise associate
with the imaging systems. Therefore, the effectiveness of convolutional neural networks
(CNN) cannot be fully achieved. In this thesis, the proposed approaches and techniques
are classified in two major categories, which are relaxation and restriction methods in
CNN models to promote the segmentation performance for medical images.
For multi-modality and multi-class brain tumor segmentation on magnetic resonance
images, the chal...[
Read more ]
Deep learning has achieved great success in different tasks for natural image data in recent
years. However, the segmentation of medical image data appears to be challenging because
of the limited manually labeled data from the medical experts, the pathological changes
and the morphological variation of the target objects, and the random noise associate
with the imaging systems. Therefore, the effectiveness of convolutional neural networks
(CNN) cannot be fully achieved. In this thesis, the proposed approaches and techniques
are classified in two major categories, which are relaxation and restriction methods in
CNN models to promote the segmentation performance for medical images.
For multi-modality and multi-class brain tumor segmentation on magnetic resonance
images, the challenges are the severe data imbalance among the different tumor sub-regions, and the great variation in terms of the tumor location, size, shape, and appearance. Therefore, to better recognize the overall tumor structure, we relax the inner
boundary constraints for tumor sub-regions. A novel loss function is proposed to automatically enforce more attention on the harder classes during training, and a symmetrical
attention module is presented to restrict the possible location of the predicted tumor.
The experimental results on the publicly available datasets from real patents validate the
effectiveness of these proposed approaches.
Colon gland instance segmentation on histological images is a crucial step for colorectal
cancer diagnosis in clinical practice, but accurate segmentation of extremely deformed
glands of highly malignant cases or some rare benign cases remains to be challenging.
Therefore, we relax the input domain to incorporate the clinical text for high-level feature guidance of the glandular objects with different histologic grades. Besides the initial
segmentation, it offers cancer grade diagnosis and the enhanced segmentation results for
full-scale assistance. In the other approach, the gland segmentation is conducted under
the restriction of hierarchical semantic feature matching from histological image pairs
in an attentive process, where both spatial details and morphological appearances can
be well preserved and balanced, especially for the glands with severe deformation or
mutation. A loss function is introduced to enforce simultaneous satisfaction of semantic
correspondence and gland instance segmentation on pixel-level. The models successfully
boost the segmentation performances on the greatly mutated or deformed cases, and
outperform the state-of-the-art approaches on the public datasets from real patients.
Post a Comment