Confounder-Aware Image Synthesis for Pathology Segmentation in New Magnetic Resonance Imaging Sequences

Jesse Phitidis, Antanas Kascenas, Maria del C. Valdés Hernández, William Whiteley, Joanna M. Wardlaw, Alison Q. O'Neil

Abstract: Given the trend for automation in radiology, MRI scan manufacturers would like to deploy their new sequences with accompanying computer aided diagnosis (CAD) algorithms. It is assumed that for each imaging sequence, we initially lack data and annotations to train medical image analysis algorithms. In this paper, we propose a method to train supervised machine learning algorithms for pathology segmentation in a specific target domain. Our method requires a dataset annotated with anatomy and pathology labels in at least one source domain, plus estimates of the intensity distribution (mean, variance) of each distinct anatomical feature/tissue type in the target domain. Intensities can be estimated using domain and MRI physics knowledge, perhaps in combination with phantom data. We then propose simplistic synthesis of training images, using the source domain dataset labels combined with additional confounder pseudo-labels for the target pathology. We test our approach on the challenging task of ischaemic stroke lesion segmentation, using T1-weighted MRI as the source domain and diffusion-weighted MRI (DWI) and fluid-attenuated inversion recovery MRI (FLAIR) as the target domains. Our method reaches 69% and 75% respectively of the Dice score of a benchmark model trained on real target domain data. The addition of the confounder pseudo-labels was effective for reducing false positives, especially for DWI scan acquisition artefacts. Code is available at: https://github.com/Jesse-Phitidis/SynthPathCode.

External IDs:dblp:conf/miua/PhitidisKHWWO24