Developing Computer Vision and Machine Learning models for analysis of Complex Data
My research focuses on dealing with core challenges of Computer Vision (CV) and applied Machine Learning (ML) to develop generalizable solutions for processing and analysis of complex data. I develop computational models to deal with high variability of image statistics, insufficient amount of labeled data, unbalanced data, and data normalization.
In this talk, I will present my bottom-up challenge-oriented approach – focusing mainly on tackling a fundamental CV / ML challenge, then proposing a novel solution and applying the developed methodology on any kind of complex data at hand, medical or natural. Applying my methodologies in the medical domain is an advantage because it encourages the development of generalizable methodologies that can handle well with high-level complex data. While the performance of current ad-hoc techniques (task-data specific techniques that are commonly used in the medical domain) decreases dramatically in real-world settings where the data is highly diverse, my broad challenge-oriented approach motivates the development of domain-invariant methodologies. I will demonstrate the strengths of the proposed solutions for relevant applications and interesting research questions. As an example, I will introduce an adaptive generalizable segmentation framework that can handle substantial diversity of image characteristics, and supply far more general, accurate and robust segmentation solution, than ad-hoc designed platforms that are currently available. I will show the capabilities of this single framework for analysis of diverse and complex dataset that includes MRI brain lesions, CT Lung nodules, CT Liver lesions and Mammography images. I will also present a novel ML architecture, based on Capsule Networks and Attention mechanism, for more accurate object classification (natural and medical), and a method for data normalization in case it contains several data distributions.