Stanford community and AIMI Affiliates only
In this talk, I will present a test-time adaptation strategy for making CNN-based segmentation algorithms robust to domain shifts that are commonly present in medical imaging and pose crucial obstacles for processing MRI. CNNs work very well for supervised learning problems as long as the training dataset is representative of the variations expected to be encountered at test time. Domain shift issues due to variations in acquisition details and scanners can violate this premise and cause remarkable performance degredation. We address this problem by prepending a shallow adaptable module to a given segmentation module and adapt the module's parameters during test time based on prior expectations of the segmented structure, encoded through a Denoising Autoencoder. Validation on multi-center MRI datasets of three anatomies: brain, heart and prostate, demonstrates the performance gains and robustness that can be achieved using the proposed test-time adaptation strategy. Being agnostic to the architecture of the segmentation CNN, the proposed design can be utilized with any network to increase robustness to variations in imaging scanners and protocols.
Dr. Konukoglu earned his B.S. and M.S. degrees at Bogazici University / Electrical and Electronics Engineering Department in 2003 and 2005. He got his PhD from University of Nice Sophia Antipolis working at INRIA Sophia Antipolis Mediterranean under the supervision of Prof. Nicholas Ayache in 2009.
After he received his PhD, he worked as a post-doctoral researcher at Microsoft Research in Cambridge between 2009 and 2012. Between 2012 and 2016 he worked at Athinoula A. Martinos Center for Biomedical Imaging and Harvard Medical School as an Instructor in Radiology and Assistant in Neuroscience. He was a member of the Laboratory for Computational Neuroimaging.
In August 2016, he started as an Assistant Professor (tenure-track) of Biomedical Image Computing at ETH-Zurich.