AUTOMATIC ROI SELECTION FOR MULTI-MODAL MEDICAL IMAGES

Abstract

The Region of Interests (RoIs) in Medical Images (MIs) contain specialized form of clinical and biological data important for medical procedures and practices. Their detection with higher accuracy and representation with more efficiency have become two important requirements of many region-based MI processing. However, these requirements are largely overlooked in contemporary literature. The development of an efficient and automatic RoI selection method has, therefore, been investigated in this thesis to attain two major requirements: (i) faster and accurate RoI segmentation and (ii) efficient representation of the segmented RoI. An Active Contour Model (ACM) based segmentation scheme is developed to effectively tackle intensity inhomogeneities and different shapes, sizes and locations of RoIs, and thereby, to automatically segment the RoIs with higher accuracy. Unlike the existing ACMs, two novel local images are constructed and fitted in the relative entropy based energy functional for better curve evolution and increasing robustness to noise and initialization. The energy equation is also scaled by local dispersion based edge mapped image for accuracy in boundary detection and smoothness. For the efficient representation of the segmented RoIs, the segmented region is defined by its original shape with reduced information by an effective polygonal decimation process. The overall performance of the proposed automatic RoI selection method is verified with the effectiveness and efficiency of the newly developed RoI segmentation and representation schemes for multi-modality MIs. Particularly, compared to the prominent and recent ACMs, the proposed segmentation scheme offers a faster curve evolution and more robustness to noisy, low-contrast, and intensity inhomogeneous MIs. The effective representation scheme is also compared with the 5 times and 10 times reduced vertices and respective bit requirements. The utilization of the proposed RoI selection method would be promising for the region-based MI processing and its security protection applications.