On the 27th of February 2024, the 40th CIBM Breakfast & Science Seminar was given by Dr. Thomas Sanchez, Research Staff Scientist, CIBM SP CHUV-UNIL, Medical Image Analysis Laboratory (MIAL). The event was chaired by Dr. Meritxel Bach Cuadra, Head, CIBM SP CHUV-UNIL Computational Neuroanatomy & Fetal Imaging Section.
Deep learning in neuroimaging: a story of domain shifts and data quality
Abstract
An increasing amount of machine learning and deep learning (DL) tools are developed for neuroimaging, promising to push further the boundaries of what is possible, from advanced data processing to innovative biomarkers and efficient computations. However, DL methods can fail when training and testing data come from different distributions (e.g. different imaging devices, hospitals, age, etc.), a case known as domain shift. But DL in neuroimaging can also suffer from poor data quality, which is known to introduce biases in downstream analyses.
In this talk, we will explore the best practices to deal with domain shifts and highlight the importance of quality control, drawing from the literature as well as our experience working on fetal brain MRI. We will show that perhaps surprisingly, sophisticated domain adaptation methods and deep learning might not always be the best solution for dealing with domain shifts in neuroimaging.
Thomas Sanchez,
Research Staff Scientist, CIBM SP CHUV-UNIL, Medical Image Analysis Laboratory (MIAL)
About the speaker
Thomas obtained from EPFL his master’s degree in computational science and engineering in 2018 as well as his PhD in Computer Science in 2022. He then joined the CIBM SP CHUV-UNIL section, headed by Dr. Meritxell Bach Cuadra, as a postdoctoral researcher. His research revolves around developing robust machine learning methods for fetal brain MRI quality control and super-resolution reconstruction, and he is interested in bringing a deeper understanding into the impact of domain shifts on neuroimaging.