LF_Deface: an Automated Tool for Defacing Pediatric Brain Low-Field MRI

Pre-processing magnetic resonance (MR) images to obscure facial features is essential for protecting subjects' privacy before publicly releasing brain image data. However, existing defacing tools designed for adult MR images struggle with pediatric data due to differences in head size, underdeveloped anatomical structures, and natural tissue contrast variation. Adaptation is further complicated while attempting defacing for Low-Field (LF, 0.064 T) MR imaging, an emerging technology that greatly benefits pediatric populations due to its portability and faster acquisition time. In this paper, we present two novel pipelines for defacing pediatric LF images. The first pipeline operates on paired High-Field (HF, >1.5 T) and Low-Field (LF) images using an nnUNet trained to deface HF images. Defaced LF images are then generated through LF-HF image co-registration. The second pipeline, unlike the first, does not require paired images. Instead, the nnUNet previously trained on HF data is directly tested on LF images that, using our previous work, have been resolution and contrast adjusted into so-called Super-Field (SF) images; defaced MR images are then generated from SF volumes. Initially, two cohorts of paired HF-LF MR images were used to train our nnUNet. After training, we verify our pipelines in a zero-shot manner across multi-site images to underscore the efficacy of our approach as the first dedicated solution for LF pediatric MR image defacing.