Bayesian Analysis of Medical Images from Weak Supervision
Improvements in data acquisition technologies make ever larger data masses available for machine learning. However, data annotation capabilities do not keep pace with the amount of available raw data. Recognition of patterns of interest requires ground-truth labels, gathering of which comes at the cost of manpower in most applications. Even worse, in some fields such as medicine, only domain experts can provide annotations. My research addresses this severe annotation scarcity problem of the present big data age from diverse perspectives, such as multiple instance learning, transfer learning, and active learning. In this talk, I will go through various solutions I developed to alleviate the annotation scarcity problem in different medical image analysis tasks using novel variants of Gaussian processes (GPs). I will show how GPs can perform state-of-the-art novelty detection from few annotations (requiring only <3% of a training set to be annotated), how they can effectively model big data (> 1M data points) based only on weak labels, and how they can build layered network architectures to form highly powerful learners. I will conclude by giving a glimpse of my current running research activities and future plans.
Melih Kandemir received his B.Sc. and M.Sc. degrees in Computer Engineering from Hacettepe University, Ankara, Turkey, in 2005 and Bilkent University, Ankara, Turkey, in 2008, respectively. He joined the Statistical Machine Learning and Bioinformatics research group of Aalto University School of Science, Espoo, Finland, in 2008 and earned his Ph.D. degree in 2013. Since 2013, he is with Heidelberg Collaboratory for Image Processing (HCI) at Heidelberg University in Germany. His main areas of interest include Bayesian modelling and inference, Gaussian processes, weakly-supervised learning, and application of all these areas to challenging real-world medical image analysis, computer vision, and neuroinformatics tasks.
December 30, 2015 – 11:00, FENS L027