Automated brain tumor segmentation and detection are immensely important in medical diagnostics because it provides information associated to anatomical structures as well as potential abnormal tissue necessary to delineate appropriate surgical planning. In this work, we propose a novel automated brain tumor segmentation technique based on multiresolution texture information that combines fractal Brownian motion (fBm) and wavelet multiresolution analysis. Our wavelet-fractal technique combines the excellent multiresolution localization property of wavelets to texture extraction of fractal. We prove the efficacy of our technique by successfully segmenting pediatric brain MR images (MRIs) from St. Jude Children’s Research Hospital. We use self-organizing map (SOM) as our clustering tool wherein we exploit both pixel intensity and multiresolution texture features to obtain segmented tumor. Our test results show that our technique successfully segments abnormal brain tissues in a set of T1 images. In the next step, we design a classifier using Feed-Forward (FF) neural network to statistically validate the presence of tumor in MRI using both the multiresolution texture and the pixel intensity features. We estimate the corresponding receiver operating curve (ROC) based on the findings of true positive fractions and false positive fractions estimated from our classifier at different threshold values. An ROC, which can be considered as a gold standard to prove the competence of a classifier, is obtained to ascertain the sensitivity and specificity of our classifier. We observe that at threshold 0.4 we achieve true positive value of 1.0 (100%) sacrificing only 0.16 (16%) false positive value for the set of 50 T1 MRI analyzed in this experiment.
The problem of automatic target recognition (ATR) and image classification have been active research fields in image processing. In this research, we explore ATR techniques such as object pre-processing, detection, tracking and classification for sequence of infrared (IR) images. The detection and tracking of IR images is performed using Bayesian probabilistic technique. The tracked part of the object frame is then processed to discard the background to obtain just the segmented object. The segmented dataset is then rendered shift invariant by first calculating the mean of the object and then moving the mean to center of the frame. We divide each frame into blocks and obtain statistical features such as mean, variance, minimum and maximum intensity in each block for subsequent classification. We visually divide entire IR dataset into 8 classes for supervised training using a K-nearest neighbor classifier. We classify the test IR dataset into 8 different classes successfully.
The automatic target recognition (ATR), often time, is limited by the presence of background clutter and distortions such as scale, translation and rotation (both in-plane and out-of-plane) in both single and multi object cases. Such distortion invariant ATR and image understanding have been the subject of intense
research in machine vision. In a previous work, we have demonstrated the usefulness of an amplitude-coupled minimum-average correlation energy (AC-MACE) filter in in-plane rotated SAR image ATR. The AC-MACE filter outperforms the regular MACE filter in rotation-related cases. Motion tracking is also an important task in computer vision, especially, when objects are subjected to certain viewing transformation. There are many problems in which very small objects undergoing motion must be detected and then tracked. For example, one of the most difficult goals of ATR is to spot incoming objects at long range, wherein the motion seems small and the signal to noise ratio (SNR) is poor. The system must be able to track such targets long enough to identify whether the object is a friend or foe. In this work, we are interested in locating both long-range and short-range moving objects in IR images wherein the object may vary from a few pixels in size to a large number of pixels in a sequence of IR images. The targets are submerged in background noise and clutter. Additionally, the tracking problem also involves out-of-plane rotation of the target. Thus, we investigate both MACE and AC-MACE filter for rotation and size invariant target detection and tracking using realistic IR images.
In the previous paper [2] we demonstrated the usefulness of complex frequency scaling of Fourier Transform in identifying amount of rotation angle between two objects. Using the complex frequency scaling property of Fourier transform, an image and its Hilbert transform can be used to find the exact angle of rotation between the images [24]. In order to find the correct angle between the two different images, we need to find the Hilbert transform of the function f(x,y) to construct an analytically extended function f'(x,y) . However, our approach does not perform satisfactorily for identification of rotation angle between two similar objects [2]. A considerable amount of research has been performed on wavelet based signal processing by utilizing a pair of wavelet transforms where the wavelets form the Hilbert transform pair. In this paper we describe the design procedure based on spectral factorization in the generation of the Hilbert transform pair of wavelet bases. [1], [3], [4], [5]. The one-dimensional wavelets are then used to generate two-dimensional wavelets. The two-dimensional wavelets thus generated are then used in the determination of correct angle between the images. The intention behind taking the above approach using the wavelets is to find if the wavelets help in discriminating the different images. In this paper we use the Hilbert transform pair of wavelet bases instead in constructing the analytically extended function. In generating the filter, orthogonal solutions are presented [3]. The solution depends on the all pass filter having a flat delay response [10]. We use the infrared images to validate our algorithm.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.