An unexpected confounder: how brain shape can be used to classify MRI scans ?
Keywords: Confounding factor, Classification, Brain shape, Deformable registration, Interpretability
Abstract: Although deep learning has proved its effectiveness in the analysis of medical images, its great ability to extract complex features makes it susceptible to base its decision on spurious confounders present in the images. However, especially for medical applications, network decisions must be based on relevant elements. Numerous confounding factors have been identified in the case of brain scans such as gender, age, MRI sites or scanners, etc. Nevertheless, although skull stripping is a classic preprocessing step for brain scans, brain shape has never been considered as a possible confounder. In this work, we show that brain shape is used in the classification of brain MRI scans from different databases, even when it should not be considered as a clinically relevant factor. To this purpose, we introduce a rigorous two steps method to assess whether a factor is a confounder or not, and we apply it to identify the brain shape as a confounding variable in brain images classification. Lastly, we propose to use a deformable registration in the data preprocessing pipeline to align the brain contours of the images in the datasets, whereas standard pipelines often do nothing more than affine registration. Including this deformable registration step makes the classification free from the brain shape confounding effect.
Submission Number: 259
Loading