PigBET: A 2.5D deep learning segmentation framework for multimodal and longitudinal domestic pig MRI utilizing ImageNet pre-trained encoders

Abstract

The domestic pig model is increasingly utilized in developmental neuroscience studies due to its anatomical and functional similarities to the human brain. Neuroimaging has become an important tool for studying pig brains. Skull Stripping is a critical but labor-intensive task for processing pig magnetic resonance imaging (MRI) data, presenting a significant hurdle in longitudinal and multimodal neuroimaging studies utilizing pig as a biomedical model. To address this, we developed PigBET, a deep learning-based skull stripping optimized for domestic pig. We stacked adjacent slices into three channels, capturing additional spatial context along the third axis. This strategy enables the use of pre-trained encoders, designed for natural images, for a 3D segmentation task. To further improve accuracy, we trained separate models for the axial, coronal, and sagittal planes, and combined their predictions using majority voting. This ensemble strategy enhances robustness and reduces orientation-specific errors. Training utilized a large quantity of 4-week-old pig magnetization-prepared rapid gradient-echo (MPRAGE) data, with transfer learning applied to 1-week-old, 8-week-old, and 18-week-old pig MPRAGE and 4-week-old pig dMRI data. PigBET achieved high segmentation accuracy on the primary test set with a mean Dice coefficient of 0.981 ± 0.004, IoU of 0.962 ± 0.008, and Hausdorff Distance (HD) of 3.4 ± 1.28 voxels. With transfer learning, it maintained robust performance across modalities and ages: Dice = 0.975–0.982, IoU = 0.952–0.965, HD = 1.7–3.5 voxels. When benchmarked against other methods, PigBET outperformed another U-Net approach without pre-trained encoders that previously showed capability of skull stripping pig data as well as a standard registration-based method combining FLIRT and ANTs. These results demonstrate that PigBET is robust and efficient and accommodates pig data from various ages and different MRI contrasts. This tool significantly advances the efficiency of MRI data processing for large-scale pig neuroimaging studies, making it a valuable resource for the biomedical field.