Filipe Condessa Seeks to Make Medical Image Segmentation Faster And More Efficient
Co-written with Filipe Condessa’s co-advisors, José Bioucas-Dias (IST-UL) and Jelena Kovacevic (CMU), the paper presents a method that makes the classification of segmented images faster and more effective. “The paper was the basis for an invited presentation I did at the IS&T/SPIE 2014 Electronic Imaging Conference,” said Filipe Condessa and “in it we present a novel algorithm for image segmentation which separates itself from the generally discrete nature of the algorithms for image segmentation as it introduces a continuous hidden field that drives the segmentation,” the student explained.
In computer vision, image segmentation is the process of partitioning a digital image into multiple segments in order to simplify and/or turn the representation of an image into something that is more meaningful and easier to analyze. According to Filipe Condessa, t he main impact of the method presented in his paper “is the ability to quickly find optimal solutions to segmentation problems by formulating them as a convex problem, instead of an integer optimization problem where optimal solutions are not attainable in a timely manner,” he clarified, adding that “this method allows us to obtain a better classification performance in large datasets.”
In digital pathology, for instance, the goal is to discover and classify tissues in images acquired by microscope imaging, and the method presented uses the contextual information more effectively, thus expediting the classification process. “Comparatively to previous methods, the prior information required is one order of magnitude lower. This makes it easier for the histopathologist to train the classifier,” explained Filipe Condessa.
The method can be applied to automatically classify tissue in digital pathology and to discover and categorize landmasses in multidimensional images acquired by remote sensing instruments (hyperspectral imaging). Here, “the results allow us to obtain state of the art results,” said Filipe Condessa, enrolled in the dual degree program since the academic year 2011/2012.
“At the moment my research is focused on improving the performance of the classifier without improving the classifier itself. We add contextual information and a rejection option to a generic classification system to improve the performance,” the student stated explaining that the method achieved an 85 percent performance on 90 percent of the data set, as opposed to the previous 55 percent achieved with traditional methods.
Image segmentation can not only be used for medical purposes (for instance, to locate tumors and other pathologies, measure tissue volumes, medical diagnosis, plan surgeries), but also in content-based image retrieval and object detection, among others.