Image Segmentation-Registration Cooperative Techniques Applied to Biomedicine and Remote Sensing.

다국어 초록 (Multilingual Abstract)

The topic of this dissertation is the development of segmentation–registration and registration–segmentation cooperative techniques for applications in remote sensing and biomedical imaging domains. On one hand, image segmentation is the process of partitioning digital images into meaningful regions, while on the other hand, image registration which is the process of establishing a correspondence for the alignment of two images of the same scene portraying viewpoint, temporal or sensor variations. Image registration and segmentation are useful in many domains of application in computer vision and biomedicine. In biomedical applications, accurate region delineation is useful for tissue identification, tissue quantification and cell tracking. In addition, correct alignment of images may be critical for detection of tumors, changes in body tissues, the effect of aging on the bone structure, or diseases in the body. Hence aiding treatment planning, clinical diagnosis, and clinical trials. In satellite–remote sensing applications, segmentation may be used for target recogniction, and object detection, while a precise alignment of images may be important for target and change detection, urban planning, and many other applications. In the part of this work that relates to automated registration via segmentation, we propose a joint region feature set and a matching cost function for registration of remote sensing/satellite images exhibiting multi-sensor, multi-spectral, and multi-view characteristics. Our joint region descriptor is a combination of Fourier, intensity, and shape features. Our work first delineates the input and reference images into regions by segmentation. Next, it extract joint intensity and shape features from the regions. It computes region similarity measures for feature matching. Finally, it estimate the geometric transform by the maximum-likelihood sampling consensus (MLESAC) technique. Our results are promising, showing potentials for sub-pixel accuracy. In addition, our algorithm outperformed most state-of-the-art automated feature-based registration algorithms incorporating feature detection and extraction algorithms such as SURF-SURF, BRISK-SURF, HARRIS-FREAK, FAST-FREAK, minimum eigenvalue-FREAK, and KAZE-KAZE.Furthermore, this work proposes an unsupervised subspace learning-based disaster mapping (SLDM) technique using pre- and post-disaster satellite imagery. It first finds the geometric transformation for automatic image registration by matching regions represented by shape and intensity descriptors as described above. It produces piece-wise constant approximations of the two images using the delineated regions. It performs subspace learning in the joint regional space to produce a change map and identify the damaged regions by probabilistic subspace distances between test points and the subspace model.We tested our method on seven disaster datasets including four wildfire events, two flooding events and one tsunami/earthquake event. Comparative analysis of our method with state-of-the-art techniques, such as Gabor Two-Level Clustering (G-TLC), image differencing and differenced-spectiral indices (NDVI, NBR, and NDWI) indicate that the proposed SLDM framework and especially the multi-band variant SLDM-SF-MS-MB produced more accurate disaster maps than the compared method, overall. Quantitative and qualitative evaluations of the proposed SLDM method confirm its capacity for disaster mapping, and its applicability to different categories of disaster events. In the segmentation approach using deformable registration models, our work first introduces a statistical shape modeling technique for atlas generation and multi-atlas-based image segmentation (MAIS) techniques for the identification of soft and hard tissues in biomedical imaging data. To calculate the deformation fields, it employed multi-grid free-form deformation (FFD) models with B-splines, symmetric log-domain extension of diffeomorphic demons (SDD) or Symmetric Normalization (SyN) from the Advanced Normalization Tools (ANTs) library. It then applies majority voting, or Simultaneous Truth And Performance Level Estimation (STAPLE) for label fusion. It compared the results of our MAIS methodology for each deformable registration model and each label fusion method, using Dice similarity coefficient scores (DSC) against manually segmented tissue label maps. It applied MAIS methods: FFD-STPL, SDD-STPL, and SyN-STPL, to two datasets: 2D peripheral quantitative computed tomography (pQCT) scans in the lower leg and 3D thigh magnetic resonance images (MRIs), where STPL stands for simultaneous truth and performance level estimation or STAPLE. Our results are encouraging for both applications, even for pQCT scans with considerable quality degradations – such as motion artifacts, or when boundaries between tissues are ambiguous because of physiological reasons. SDD-STPL outperformed FFD-STPL and SyN-STPL on both datasets. Furthermore, our method applied the BM3D algorithm to enhance the statistical atlas generated for the 3D thigh MRI dataset, and generated silver truth with our MAIS model. Statistical enhancement using BM3D improved our MAIS result better than other image enhancement techniques, including Wiener filtering, unsharp-masking, histogram equalization and histogram matching.