Introduction: Tetramethylenedisulfotetramine (TETS), a neurotoxic rodenticide, triggers seizures and status epilepticus (SE) in humans by antagonizing the γ-aminobutyric acid type A (GABA_A) receptor [1]. No specific medical countermeasures exist for TETS-induced SE [2], driving significant interest in using MRI and PET in mouse models of acute TETS intoxication with atlas-based analysis to understand neuropathology and assess novel treatments. Whole brain delineation (WBD) improves atlas-based segmentation by eliminating unwanted signals through skull stripping [3]. Recently, we developed a modified 2D U-Net for WBD in rat brain MRI [4] of an organophosphate intoxication (OPI) model [5], which we directly applied to brain MRIs of the TETS mouse model. This method resulted in low accuracy segmentations in regions with low signal-to-noise ratio and erroneously included the ear canals. [6]. To overcome these challenges, we created an OPI-pretrained 3D U-Net [7] that improved segmentation accuracy but was still limited. In this work, we investigated if transfer learning (TL) with the OPI-pretrained 3D U-Net would enhance WBD accuracy in brain MRIs of the TETS mouse model (1) compared to direct application, and (2) a TETS-only 3D U-Net.

Methods: Our data consisted of T₂-weighted MRI rat (n=100 scans) and mouse (n=85) brain images acquired on a Bruker BioSpec 7T scanner (phased array coil, voxel size: 125x125x500μm, matrix size: 280x200x59 and 160x160x35, respectively). These data consisted of scans of adult Sprague Dawley rats from an OPI study [8], and adult C57BL/6J mice from a TETS study [9]. OPI rats were organized into 5 groups: vehicle controls (VEH), untreated OPI animals (DFP), and DFP animals treated with midazolam (MDZ), allopregnanolone (ALO), or combined MDZ and ALO (DUO). For the TETS study, mice groups were divided into VEH, MDZ, ALO, and DUO. Each scan was segmented by an experienced human observer to delineate the whole brain, excluding the olfactory bulb and brainstem. The 3D U-Net was modified to have 18 convolutional layers (activation function=leaky ReLU), three max pooling layers with zero padding, and one dropout layer to prevent overfitting [7]. The OPI rat U-Net was trained over 300 epochs with 100 scans (DFP=23, MDZ=23, ALO=21, DUO=21, VEH=12). The TETS mice-pretrained U-Net and OPI rat U-Net with TL, were trained over 200 epochs, where the training and evaluation datasets consisted of 73 scans (MDZ=30, ALO=14, DUO=20, VEH=9) and 12 scans (MDZ=3, ALO=3, DUO=3, VEH=3), respectively. Image and label data were preprocessed, and underwent data augmentation during training as previously reported [6]. Parametric selection of the training dataset size (TDS) for OPI rat-pretrained U-Net with TL was evaluated by test multiple sets [TDS:1-50] and the best model was chosen for comparison. The Dice coefficient (DC) and Hausdorff distance (HD) were calculated between the U-Net-generated labels and manual segmentations to compare accuracy between networks.

Results and Conclusion: The OPI rat-pretrained U-Net achieved a DC (median[range]) of 0.956 [0.945-0.970] and HD of 1.10[0.74-1.95]mm. These metrics were improved for the OPI rat-pretrained U-Net with TL, which achieved a DC of 0.992[0.984-0.993] and a HD of 0.42[0.25-0.53]mm. The TETS mice-pretrained U-Net achieved a DC of 0.991[0.983-0.993] and an HD of 0.40[0.31-0.50]mm, comparable to that for the OPI rat-pretrained U-Net with TL. A TDS=10 achieved the lowest median HD for TL without significantly changing DC. Thus, we conclude that the OPI rat-pretrained U-Net with TL provided a fully automated and efficient segmentation method for WBD in mice with only a small training dataset to achieve a similar accuracy to the TETS mice-pretrained network. Future research will focus on exploring transfer learning as a means for brain segmentation in other animal models.

Image/Figure:
Click to View

Image/Figure Caption:

Figure: A) Architecture of the 3D U-Net utilized. Each blue box is a feature map with multiple channels, where the number of channels is denoted above the box, and the x-, y-, and z-dimensions are denoted in the lower left of the first box in each multi-layer row. The light blue boxes represent concatenated feature maps from previous layers. The arrows indicate different operations in the neural network.  B) Boxplots (error bar denote 1.5 IQR) of (Top) Dice coefficients (DC) and (Bottom) Hausdorff distances (HD) at different training dataset sizes (TDS). Median, first and third interquartile ranges are shown. C) Table shows Dice coefficients (DC) and Hausdorff distances (HD) by group and by timepoint for each of the three networks: (left to right) OPI rat-pretrained 3D U-Net, OPI rat-pretrained 3D U-Net with TL at TDS=10, and TETS mice-pretrained 3D U-Net. D) Images from two different evaluation scans of that were segmented with the 3D U-Net with TL at TDS=10. The first row shows three views of the T2 weighted MRI of each animal (left to right: axial, coronal, and sagittal). The second row shows the U-Net-generated label on the anatomical image with the same three views. The green voxels show correctly segmented voxels while the red voxels indicate the incorrectly labeled voxels. The ALO Day 7 scan achieved the lowest HD and second highest DC, while the VEH Day 7 scan achieved the highest HD, and second lowest DC. The VEH Day 7 scan shows that a limitation of the U-Net is that it can under-segment the anterior edges (or slices) of the brain.

Author