SAMson: an automated brain extraction tool for rodents using SAM

Abstract

Accurate brain extraction is a critical step in the analysis of rodent head magnetic resonance imaging (MRI) data. However, current methods often encounter difficulties in handling the diverse range of imaging setups, resolutions, and experimental conditions that are commonly found in this field. Based on the Segment Anything Model (SAM), we introduce here SAMson (SAM for Segmentation Of Neuroimages), an automated tool for robust rodent brain extraction. SAMson integrates a bounding box generator and a mask prediction pipeline, offering fully automated and semi-automated modes to address varying experimental complexities. The performance of SAMson was evaluated using three multi-centre rodent MRI datasets annotated at the pixel level, which differed in terms of acquisition parameters, resolution, and animal age groups. SAMson demonstrates superior performance to existing methods, including BET, RBM, and BEN, in terms of segmentation accuracy, with Jaccard indices exceeding 90% across datasets. The semi-automated mode demonstrates particular efficacy in challenging scenarios, including low-resolution images and cases requiring refined mask precision. In contrast to conventional volumetric techniques, SAMson identifies errors at the level of individual slices, thereby enabling rapid and targeted correction when needed. By providing open-source access, SAMson aims to support large-scale research workflows and advance translational neuroscience. The curated data can be downloaded from https://doi.org/10.20350/digitalCSIC/17000, and the code is available at https://github.com/CanalsLab/SAMson.