Laboratories

Machine Learning

Research Outline

Machine Learning for Biomedical Imaging
Top row: Bayesian eigenspectra MSOT inversion grid and corresponding graphical models. Bottom row: sO2 values obtained from the background and vessel areas by three algorithms (linear unmixing, eigenspectra MSOT and Bayesian eigenspectra MSOT) depending on SNR of the optoacoustic images
Figure: CBI

We draw upon machine learning (ML) theory to solve problems and create new knowledge in medicine and biology by the way of furthering biomedical imaging. We target developing new ML analytic tools suited to the specific nature of biomedical images.

Imaging biomedical phenomena and processes renders precious and unique data which do not directly fit the commonplace methodology of machine learning. Specifically, one can make three observations:

The first observation is that the data itself differs in content and nature, from the kind of data that common machine learning and deep learning algorithms and architectures are designed or pre-trained for (i.e. natural images). Tissue images come often in multi-dimensional format with unique distortions and noise characteristics and often reveal features and objects which diverge from components of natural images. The second observation relates to the relative scarcity of biomedical data. Labeling data is extremely expensive in medicine and biology. The third observation is about the need for "accountable algorithms" in the general domain of healthcare. The black-box nature of common ML architectures might hinder their application, even given their accuracy, in the healthcare domain.

Given the above three observations, our principal mission is to develop data-efficient (i.e. trainable with few labels) and accountable algorithms for biomedical imaging, with diagnostic, predictive and therapeutic applications.

We seek talented Ph.D. students and postdocs with theoretical and programming competence as well enthusiastic collaboration partners, sharing a passion for the principal mission stated above.

Relevant publications

1. P. Mohajerani, V Ntziachristos, "An inversion scheme for hybrid fluorescence molecular tomography using a fuzzy inference system", IEEE Transactions on Medical Imaging 35 (2), 381-390.

2. I. Olefir, et. al, "A Bayesian Approach to Eigenspectra Optoacoustic Tomography", IEEE Transactions on Medical Imaging.

3. Q. Mustafa,et. al, "Fast Three-dimensional Weighted Optoacoustic Reconstruction using the Omega-k Algorithm", in preparation.

4. P. Mohajerani, et. al, "Fourier Backprojection for Frequency Domain Optoacoustic Tomography: Analysis and Experimental Results", in preparation.