Skip to main content

Statistically Efficient Structured Prediction for Computer Vision and Medical Imaging

Final Report Summary - STRATEGIE (Statistically Efficient Structured Prediction for Computer Vision and Medical Imaging)

This project involves the development of statistical methodologies for structured prediction and their application to computer vision and medical image analysis. Structured prediction is ubiquitous in computer vision, such as in segmentation and object localization.
Advanced imaging techniques have revolutionized the medical profession, allowing medical professionals unprecedented access to visualizations of patients internal organs and tissues without invasive procedures. At the same time, this has lead to an explosion of data that is available. A medical professional using computer tomography may now need to examine many hundreds of visual slices per patient, and full body scans are feasible. In this setting, automated tools are necessary to process the images, and to draw focus to anomalous and important regions of the data. Statistical methods are essential to address these issues due to imperfections in imaging sensors, physical limitations of imaging technologies, and variation in the human population. In particular, we will make use of structured prediction, a general framework for statistical learning and inference when predictions have complex structure or interdependencies. This setting works by providing a training set of labeled examples of medical images from which a learning algorithm tunes parameters of an automated algorithm to optimize the expected performance of the system. In this way, partially automated image analysis systems may achieve improved performance at reduced cost to the reliance on human experts.

Although the primary application considered in this work is medical imaging, both for its impact and as a model domain, structured prediction methods are more widely applicable. They are found in such diverse domains as search optimization, machine translation, drug discovery, RNA secondary structure prediction, and computer vision. This project will develop core structured prediction technologies that will have impacts in these and other as yet undefined application areas.
Project Objectives
Professionally, the main objectives of the reporting period are to build a research team capable of delivering top quality scientific publications at top venues in machine learning, computer vision, and medical image analysis. Also, a key objective is to place myself with a professorship at a top ranked European institution at which I can have long term stability and develop a strategy to maximize long term scientific impact.
The scientific objectives for the reporting period involve the development of mathematical results, algorithms, and prototype systems advancing the state of the art in structured prediction for medical image analysis and computer vision. This has been approached primarily by a strong focus on the marriage of graphical models with discriminative statistical learning techniques.
The main methodological objectives are the development of segmentation models, learning with non-modular losses, hypothesis testing in graphical models, and the development of methods for structured sparsity regularization - an important tool for learning when data are expensive and/or rare as is often the case in medical image analysis.
Applications to computer vision and medical image analysis: A diverse sample of applications are essential to include in the research agenda as a result of the no free lunch theorem of machine learning which states that no single learning algorithm will obtain the best possible performance across all problem domains. It is therefore not possible to develop machine learning algorithms in isolation, but empirical evaluation on challenging real-world tasks is essential.

Work progress during the project:
During the project, I have obtained a professorship at KU Leuven, a top European university, where I lead an active research team at the interface of machine learning, computer vision, and medical image analysis. I have developed collaborations in all three areas, leading to a large number of high impact publications. Four students have completed their PhDs under my supervision during the course of the project. The project webpage with links to open access versions of all publications, as well as links to contact involved researchers is