Uncertainty quantification
Building trust in Artificial Intelligence (AI) is a key for bringing machine learning - based solutions into clinical practice. Our division therefore puts a strong focus on the quantification and compensation of uncertainties related to the proposed image analysis methods. While we have a long history of uncertainty handling in general (see e.g. (Maier-Hein et al., 2012; Maier-Hein et al., 2016; Moccia et al. 2018)) our current research is directed to out-of-distribution (OoD) detection and the systematic handling of ambiguities in inverse problems.
© dkfz.de
Out-of-distribution detection
A common criticism of machine learning-based solutions is the way anomalies are handled. If a measurement is "out-of-distribution" (meaning that it does not resemble the training data), the algorithm cannot make a meaningful inference, and the probability of failure (error) is high. In our research, we address this epistemic uncertainty with an information theoretic approach based on the widely applicable information criterion (WAIC) (Adler, Ardizzone, Ayala et al. 2019; Adler, Ayala et al. 2019). This approach heavily relies upon recent methodology related to invertible neural networks (INNs) (Ardizzone et al., 2019) as their tractable Jacobian is what allows us to compute WAIC.
Ambiguity in inverse problems
While a lot of research has been dedicated to addressing uncertainty related to the potential intrinsic randomness of the data generation process (aleatoric uncertainty) as well as to insufficient training data (epistemic uncertainty), a type of uncertainty that has received very little attention in the literature is the potential inherent ambiguity of the problem. The field of spectral imaging, for example, has put a research focus onto converting high-dimensional multispectral measurements to underlying clinically relevant tissue properties, such as tissue oxygenation (see (Wirkert et al., 2016; Wirkert et al., 2017) and Fig. 1). However, state-of-the-art approaches to multispectral image interpretation typically provide point estimates and neglect the fact that the problem may be inherently ambiguous. Consequently, the estimations cannot generally be trusted to be close to the ground truth. In a joint project with the Heidelberg Institute for Scientific Computing (IWR), we addressed this problem by applying invertible neural networks (INNs) (Ardizzone et al. 2019; Adler, Ardizzone, Vemuri et al. 2019) which are powerful enough to generate full probability distributions for the predicted value compared to point estimates. We have shown that for a small number of bands these probability distributions become increasingly ambiguous to the point where they start to develop so called multiple modes which makes an inversion of this problem ill-posed and so nigh impossible. In recent studies, we have generalized our method to other predominant inverse problems of the medical domain, namely photoacoustic tomography (PAT; see (Nölke et al., 2021)) and 2D-3D image registration (see (Trofimova et al., 2020)). In the future, we aim to automate the detection of modes in the probability distribution to maximize the benefit for downstream tasks.
Key collaborators
- Prof. Dr. Carsten Rother, Visual Learning Lab, Heidelberg University, Germany
- Prof. Dr. Ullrich Köthe, Visual Learning Lab, Heidelberg University, Germany
- Prof. Dr. Dogu Teber, Urological Clinic, Municipal Clinic Karlsruhe, Germany
- Prof. Dr. Beat P. Müller-Stich, Section of Minimally-invasive Surgery of the Department of General Surgery, Heidelberg University, Germany
- Dr. Hannes G. Kenngott, Section of Minimally-invasive Surgery of the Department of General Surgery, Heidelberg University, Germany
- Dr. Edgar Santos, University Clinic for Neurosurgery Heidelberg, University Clinic Heidelberg, Germany
- Dr. Janek Gröhl, VISION Laboratory, Cambridge, UK
Publications
Adler, T. J., Ardizzone, L., Ayala, L., Gröhl, J., Wirkert, S. J., Müller-Stich, B. P., Rother, C., Köthe, U., & Maier-Hein, L. (2019, April 17). Uncertainty handling in intra-operative multispectral imaging with invertible neural networks. International Conference on Medical Imaging with Deep Learning -- Extended Abstract Track. https://openreview.net/forum?id=Byx9RUONcE
Adler, T. J., Ardizzone, L., Vemuri, A., Ayala, L., Gröhl, J., Kirchner, T., Wirkert, S., Kruse, J., Rother, C., Köthe, U., & Maier-Hein, L. (2019). Uncertainty-aware performance assessment of optical imaging modalities with invertible neural networks. International Journal of Computer Assisted Radiology and Surgery, 14(6), 997–1007. https://doi.org/10.1007/s11548-019-01939-9
Adler, T. J., Ayala, L., Ardizzone, L., Kenngott, H. G., Vemuri, A., Müller-Stich, B. P., Rother, C., Köthe, U., & Maier-Hein, L. (2019). Out of Distribution Detection for Intra-operative Functional Imaging. In H. Greenspan, R. Tanno, M. Erdt, T. Arbel, C. Baumgartner, A. Dalca, C. H. Sudre, W. M. Wells, K. Drechsler, M. G. Linguraru, C. Oyarzun Laura, R. Shekhar, S. Wesarg, & M. Á. González Ballester (Eds.), Uncertainty for Safe Utilization of Machine Learning in Medical Imaging and Clinical Image-Based Procedures (pp. 75–82). Springer International Publishing. https://doi.org/10.1007/978-3-030-32689-0_8
Ardizzone, L., Kruse, J., Wirkert, S., Rahner, D., Pellegrini, E. W., Klessen, R. S., Maier-Hein, L., Rother, C., & Köthe, U. (2019). Analyzing Inverse Problems with Invertible Neural Networks. International Conference on Learning Representations. https://openreview.net/forum?id=rJed6j0cKX
Gröhl, J., Kirchner, T., Adler, T., & Maier-Hein, L. (2018). Confidence Estimation for Machine Learning-Based Quantitative Photoacoustics. Journal of Imaging, 4(12), 147. https://doi.org/10.3390/jimaging4120147
Maier-Hein, L., Ross, T., Gröhl, J., Glocker, B., Bodenstedt, S., Stock, C., Heim, E., Götz, M., Wirkert, S., Kenngott, H., Speidel, S., & Maier-Hein, K. (2016). Crowd-Algorithm Collaboration for Large-Scale Endoscopic Image Annotation with Confidence. In S. Ourselin, L. Joskowicz, M. R. Sabuncu, G. Unal, & W. Wells (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016 (pp. 616–623). Springer International Publishing. https://doi.org/10.1007/978-3-319-46723-8_71
Maier-Hein, Lena, Franz, A. M., dos Santos, T. R., Schmidt, M., Fangerau, M., Meinzer, H.-P., & Fitzpatrick, J. M. (2012). Convergent Iterative Closest-Point Algorithm to Accomodate Anisotropic and Inhomogenous Localization Error. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(8), 1520–1532. https://doi.org/10.1109/TPAMI.2011.248
Moccia, S., Wirkert, S. J., Kenngott, H., Vemuri, A. S., Apitz, M., Mayer, B., De Momi, E., Mattos, L. S., & Maier-Hein, L. (2018). Uncertainty-Aware Organ Classification for Surgical Data Science Applications in Laparoscopy. IEEE Transactions on Biomedical Engineering, 65(11), 2649–2659. https://doi.org/10.1109/TBME.2018.2813015
Nölke, J.-H., Adler, T.J., Gröhl, J., Kirchner, T., Ardizzone, L., Rother, C., Köthe, U., Maier-Hein, L. (2021). Invertible Neural Networks for Uncertainty Quantification in Photoacoustic Imaging. In: Palm C., Deserno T.M., Handels H., Maier A., Maier-Hein K., Tolxdorff T. (eds) Bildverarbeitung für die Medizin 2021. Informatik aktuell. Springer Vieweg, Wiesbaden. https://doi.org/10.1007/978-3-658-33198-6_80
Trofimova, D., Adler, T.J., Kausch, L., Ardizzone, L., Maier-Hein, K., Köthe, U., Rother, C., Maier-Hein, L. (2020). Representing Ambiguity in Registration Problems with Conditional Invertible Neural Networks. https://arxiv.org/abs/2012.08195
Wirkert, S. J., Kenngott, H., Mayer, B., Mietkowski, P., Wagner, M., Sauer, P., Clancy, N. T., Elson, D. S., & Maier-Hein, L. (2016). Robust near real-time estimation of physiological parameters from megapixel multispectral images with inverse Monte Carlo and random forest regression. International Journal of Computer Assisted Radiology and Surgery, 11(6), 909–917. https://doi.org/10.1007/s11548-016-1376-5
Wirkert, S. J., Vemuri, A. S., Kenngott, H. G., Moccia, S., Götz, M., Mayer, B. F. B., Maier-Hein, K. H., Elson, D. S., & Maier-Hein, L. (2017). Physiological Parameter Estimation from Multispectral Images Unleashed. In M. Descoteaux, L. Maier-Hein, A. Franz, P. Jannin, D. L. Collins, & S. Duchesne (Eds.), Medical Image Computing and Computer Assisted Intervention − MICCAI 2017 (pp. 134–141). Springer International Publishing. https://doi.org/10.1007/978-3-319-66179-7_16