Computer Vision Applications: Target Detection and Medical Images

Table of Contents


Computer vision applications include steps as image processing, image analysis and classification, so that decision making based on images is feasible. We develop and apply many computer vision tools to medical micrographies and synthetic aperture radar images.

1. Target detection: SAR image

Fig.1.Front propagation to segment target from a 8-looks digital image.

In collaboration with Prof. Regis Marques and Prof. Fatima Medeiros , we have designed a new framework for point target detection in synthetic aperture radar (SAR) images. We focus on the task of locating reflective small regions using a level set based algorithm. Unlike most of the approaches in image segmentation, we address an algorithm which incorporates speckle statistics instead of empirical parameters and also discards speckle filtering. The curve evolves according to speckle statistics, initially propagating with a maximum upward velocity in homogeneous areas. Our approach is validated by a series of tests on synthetic and real SAR images and compared with three other segmentation algorithms, demonstrating that it configures a novel and efficient method for target detection purpose. This research includes image analysis; object detection; partial differential equations; radar target recognition; speckle; synthetic aperture radar.

Fig.2.Level set evolution driven by speckle statistics in synthetic aperture radar image: (a) propagation expansion speed function of a synthetic image with 8-looks speckle statistics, (b) arbitrary initial level set, (c) intermediary stage and (d) the final result.

2. Breast Cancer

Fig.3. Visualization of NuMA-stained nuclei in three dimensions. Images rendered at LBNL's supercomputing facility NERSC, demonstrate our ability to segment confocal microscopy images, containing NuMA (a) into individual nuclei (b) and reveal the underlying protein organization (c).

We are developing image analysis strategies to aid early cancer detection. Previously, using high resolution fluorescence images of cultured model human mammary epithelial tissue, we showed that the radial nuclear distribution of the nuclear mitotic apparatus protein (NuMA) correlates with the phenotype of the cells (Knowles et al., 2006). Here we are testing whether texture analysis of NuMA organization can be used to classify non-malignant and malignant epithelial cells.
   Three dimensional images of cultured mammary tissue (Fig.3-4) are first normalized to correct for brightness loss due to optical penetration and to reveal local bright features of NuMA staining (Fig.3). Total DNA, stained by DAPI, was used to provide a nuclear segmentation mask which allowed textural analysis on a per cell basis (Fig.4).


Fig.4. Visualization of DAPI-stained nuclei in three dimensions. Demonstration of previous work (Knowles et al., 2006) on using DAPI-stained cells to segment individual nuclei (a) and representation of the underlying chromatin structure (b), which can be potentially explored.

Haralick texture features of the NuMA pattern were then calculated from the gray level co-occurrence matrices. Figure 6 shows a parallel coordinate plot of the results of multiple features, many of which show clear separation between proliferating non-malignant cells (Fig.6 blue), differentiated non-malignant cells (Fig.6 green), and malignant cells (Fig.6 red).
   This work shows that textural analysis of the organization of specific nuclear proteins can be used to classify human mammary epithelial cells of different phenotypes. The joint probability distribution of the NuMA fluorescence pattern was then calculated as gray level co-occurrence matrices (GLCM). Notice there is more color variation ("color spread") in Figure 5 (a) than in (b), where you can see a red dominant peak, with a narrow surrounding area - the biological interpretation resides in the fact that non-malignant cells presents high NuMA variations while these are much more homogeneous in malignant cells.


Fig.5. Gray level co-ocurrence matrix reveal spatial correlations of intensity variation within an image : joint probability distribution of 2 pixels for: (a) non-malignant and (b) malignant cell cell, with more variation in the graylevels in (a) than in (b).

Fig.6. Parallel coordinates of attributes extracted from GLCM of hundreds of (red) malignant, (blue) proliferating non-malignant and (green) mature non-malignant cells.

3. Ocular Fundus

Fig.6.Ocular fundus and vessel segmentation: (a) Segmentation result for Wavelet (b) Result of candidate extraction.

Images of the ocular fundus presents indications about retinal, ophthalmic, and even systemic diseases such as diabetes, hypertension, and arteriosclerosis. Microaneurysms are the earliest sign of diabetic retinopathy and we aim at developing robust detection of microaneurysms in digital color fundus image as an auxiliar screening procedure. The current research involves the evaluation of a set of tools for vessel segmentation and how this task influences microaneurysms detection. We also propose a new approach to detect microaneurysms using mathematical morphology. Also, we have designed a pipeline for segmentation and the feature extraction to detect candidate microaneurysms. We show that the candidate microaneurysms detected with the proposed methodology have been successfully classified by a MLP neural network (correct classification of 84%). This work is a collaboration with LabVis group.

Future Developments

1. Extend the research on target detection to oil spill and hurricane detection. 2. Detect and quantity different cell/tissue phenotypes in epigenetic abnormalities;
3. Run experiments in larger datasets to detect microaneurysms using mathematical morphology;