PIGSKIN: PIG SKull stripping with synthetic Images and Neural networks

Abstract

Robust brain extraction is an essential first step for quantitative neuroimaging, yet no reliable method exists for porcine MRI despite its importance as a translational neuroscientific animal model, particularly gyrencephalic traumatic brain injury (TBI). Porcine anatomy presents unique challenges, with extensive extracranial fat and complex textures around the olfactory bulbs that confound existing human-based tools. We present PIGSKIN, a deep learning framework trained entirely on synthetic data generated from a small set of expert annotations. Unlike conventional approaches, PIGSKIN models brain and non-brain regions separately, applying distinct clustering and transformation parameters to each. This strategy constrains variability in brain anatomy while allowing greater diversity in extracranial tissues, ensuring anatomically consistent label maps in which the brain remains embedded within its surroundings. Additional transformations introduce further spatial and intensity variability, producing a rich stream of synthetic training pairs. During inference, PIGSKIN operates in a single step at native resolution within a standardized cube, preserving fine-scale anatomical detail and enabling generalization across cohorts differing in breed, acquisition protocol, and injury model. The model achieved comparable to expert consistency accuracy (Dice ≈ 0.97), approaching the range of reported inter-rater reliability. Finally, we show that incorporating co-registered T1- and T2-weighted inputs significantly outperforms single-modality training, underscoring the value of multimodal integration for synthetic data generation. Together, these results establish PIGSKIN as the first systematically validated solution for porcine brain extraction and a framework adaptable to other large animal brain extraction tasks.

2. Introduction

The pig has historically been a valuable biomedical research model, and more recently has been utilized as a neuroscientific pre-clinical model, including advanced neuroimaging. Porcine brain injury models are increasingly recognized as valuable for neurotrauma research due to their neuroanatomical and physiological similarities to the human brain^1,2. Unlike rodent brains, porcine brains are gyrencephalic and possess higher white-to-gray matter ratios, features that make them neuroanatomically closer to humans than rodents. As such, they represent a valuable model for studying traumatic brain injury (TBI) and its structural and functional sequelae ^1-3. Magnetic resonance imaging (MRI) of pigs, before and after injury, provides a non-invasive, longitudinal window into the evolution of brain pathology and may offer critical insights into the mechanistic underpinnings of TBI-related imaging changes observed in humans.

Quantitative analysis of human MRI has consistently demonstrated sensitivity to structural and microstructural changes following TBI ^4,5. To enable meaningful cross-species comparisons and translational insights, it is critical to adapt and apply similar analytic frameworks to porcine MRI. The field of quantitative neuroimaging relies heavily on automated image processing pipelines that begin with brain extraction (i.e., skull stripping), followed by tissue segmentation, atlas-based parcellation, and the derivation of regional quantitative metrics such as volumetric and microstructural integrity measures ^6,7. While robust and well-validated tools exist for human brain MRI, including atlas-based and deep learning–based methods such as BET, HD-BET, and SynthStrip, developed for clinical and research use in large-scale datasets, they fail when applied directly to non-human species, particularly pigs ^8-10.

Translating methods of brain extraction from humans to pigs is non-trivial due to confounding anatomical differences. In contrast to the relatively clean and symmetric separation of the human brain from the skull, the porcine brain is tightly encased in heterogeneous non-brain tissues including muscle, fat, and connective tissue that obscure its boundaries on MRI (see left panel of Fig. 1). In addition, features such as a prominent olfactory bulb, the inferior extension of the pituitary region, and the frequent inclusion of the upper cervical spinal cord within the imaging field-of-view create topological complexities absent in human brains. These structural protrusions violate the assumptions of smoothness and compactness that underlie conventional brain extraction algorithms ^11,12, leading to unreliable or failed segmentation when applied to pig imaging data (see right panel of Fig. 1). Addressing these challenges is critical for enabling automated, scalable, and reproducible analysis of porcine MRI in translational research. Despite the critical need for automated brain extraction in pigs, existing animal pipelines have focused primarily on rodents ¹³, ovine ¹⁴, non-human primates¹⁵, or neonatal piglets^11,16. While promising, these efforts demonstrate the feasibility of species-specific pipelines but remain limited in scope and fail to generalize across acquisition protocols, breeds, or injury conditions. More recently, several studies have begun to address porcine brain extraction directly; however, these methods are still restricted to neonatal cohorts, heavily manual annotations, or single-modality training, and no robust solution yet exists for adult porcine MRI across modalities and injury models. As such, manual delineation remains the de facto standard, which is time-consuming, labor-intensive, and requires specialized anatomical knowledge of the porcine brain that is scarce. Moreover, manual annotation is not scalable for high-throughput analyses or for studies involving longitudinal data.

Figure 1.

Comparison of brain extraction performance in human versus porcine MRI using a human brain extraction tool. A) Axial T1-weighted images of a human (left) and pig (right) brain. B) Results of FSL BET applied to both images (red mask). In the human case, the extracted mask closely follows brain boundaries, while in the pig brain, numerous errors are observed: false positives (FP; extracranial tissues incorrectly included as brain) and false negatives (FN; true brain regions incorrectly excluded). These failures arise because the pig brain does not follow the smooth, rounded morphology assumed by human-based tools and is surrounded by complex extracranial structures, particularly around the olfactory bulbs and fat layers. This figure highlights the need for a customized porcine brain extractor that accounts for species-specific skull–brain morphology and motivates the development of PIGSKIN.

We developed PIGSKIN (Porcine Brain Extraction Tool), the first automated framework for pig brain extraction in MRI using fully synthetic data. Our approach combines a single-stage U-Net with a dynamic synthetic data generator that models foreground (brain) and background separately, expanding limited annotations into diverse, anatomically realistic image–mask pairs. By leveraging multimodal clustering of T1- and T2-weighted images, PIGSKIN achieves comparable to expert consistency performance and robust generalization across modalities, injury models, and datasets. This establishes PIGSKIN as the first robust solution for adult porcine brain extraction and a broadly applicable framework for annotation-scarce neuroimaging domains.

3. Materials and Methods

3.1. Overview

We developed a deep learning pipeline for automated porcine brain extraction in T1- and T2-weighted MRI to address the lack of species-specific tools and the inability of human-focused methods to account for the distinctive brain morphology and complex extracranial structures of pigs. As a first step, we introduced a multi-channel clustering framework to generate initial label maps. To our knowledge, this is the first clustering approach for porcine MRI that leverages both T1- and T2-weighted contrasts, where extracranial tissues such as fat layers and the olfactory bulbs become more conspicuous, producing intricate regions that often confound conventional algorithms. By explicitly capturing these features, the resulting label maps provide greater generalization than those derived from single-modality clustering. Building on this foundation, the clustered label maps were used as the substrate for synthetic augmentation: random transformations such as scaling and rotation were applied, and because the entire brain moves consistently with the deformation, its identity is preserved. When paired with intensity-varied images, this process yields a continuous stream of realistic image–mask pairs, enabling stable training until convergence. Importantly, once trained, the model weights will be released and can be applied directly to new datasets without further training, unless fine-tuning is required for a specific modality. Together, this pipeline establishes the first robust, scalable, and readily deployable solution for adult porcine brain extraction.

3.2. Datasets

To ensure generalizability, we used data from two independent porcine traumatic brain injury (TBI) cohorts representing different injury mechanisms (focal and diffuse) and pig strains (Yucatan and Hanford minipigs). The first cohort employed the Controlled Cortical Impact (CCI) model, a well-established paradigm of focal injury. Five pigs were scanned longitudinally at baseline (pre-injury) and at multiple post-injury time points (3 days, 1 month, 3 months, and 6 months), yielding 13 imaging sessions in total. The second cohort used a rotational acceleration model, which produces diffuse axonal injury without focal lesions. This cohort also included five pigs scanned at pre- and post-injury time points, yielding 10 imaging sessions in total. Together, these cohorts provided complementary datasets spanning distinct types of brain pathology.

For the CCI cohort, structural MRI volumes were acquired at approximately 0.54 mm isotropic resolution (matrix ≈ 280 × 192 × 352). Both T1- and T2-weighted scans were obtained, and T2 volumes were rigidly registered to the corresponding T1 images for alignment. For the rotational acceleration cohort, T1-weighted images were acquired with a matrix of 192 × 256 × 160 and voxel size of 0.70 mm isotropic; only T1-weighted data were available for this cohort. All images were resampled to 0.54 mm isotropic resolution for consistency across datasets and to enable high-resolution segmentation. Downsampled versions at 1.0 mm isotropic resolution were additionally generated for the coarse localization step of PIGSKIN. Acquisition parameters and imaging characteristics for both cohorts are summarized in Table 1.

Table 1. Dataset characteristics and acquisition parameters.

Details of the experimental datasets used in this study, including MRI sequence type, acquisition parameters (TR, TE, TI, flip angle), matrix dimensions, and voxel size. Both T1- and T2-weighted scans were acquired for the controlled cortical impact (CCI; focal injury) cohort, whereas only T1-weighted scans were available for the rotational (diffuse injury) cohort. Acquisition parameters were extracted from scanner metadata, and voxel dimensions were derived from the NIfTI headers.

Dataset	Sequence	TR (ms)	TE (ms)	TI (ms)	FA (°)	Matrix size	Voxel size (mm)
CCI (focal)	T1-weighted	2080	4	1050	15	280 × 192 × 352	0.54 isotropic
CCI (focal)	T2-weighted	3200	405	N/A	120	280 × 192 × 352	0.54 isotropic
Rotational (diffuse)	T1-weighted	1770	3.8	950	9	192 × 256 × 160	0.70 isotropic

Data annotation:

Manual brain annotations were performed by two experts using ITK-SNAP ¹⁷.Expert 1 has over 10 years of experience and Expert 2 has over 3 years of experience. Each expert independently used the 3D round-shaped paintbrush tool to delineate the brain’s soft tissue on axial slices, proceeding slice by slice. The inferior boundary of the annotation was defined at the lowest portion of the medulla. Each subject was inspected across three orthogonal planes (coronal, axial, sagittal) to ensure smooth boundaries and 3D anatomical consistency. For illustration, we report annotation details on a representative baseline subject (animal ID 81, pre-injury scan; hereafter referred to as the reference subject). For this subject, annotation required approximately one hour, involving review of ~170 coronal slices, ~100 axial slices, and ~100 sagittal slices. On average, annotation required ~10 seconds per slice for Expert 1 and ~15 seconds per slice for Expert 2, with additional time devoted to challenging regions such as the olfactory bulbs and brainstem to ensure accurate delineation. When disagreements arose, annotations were jointly reviewed and resolved through a consensus adjudication process, yielding a finalized reference mask used for all downstream training and evaluation. Most ambiguities were localized to the brainstem, particularly at its inferior boundary, and were resolved through an additional focused review and correction pass. The level of agreement between the experts was quantified.

Training/Test Split

For the CCI cohort, seven unique pigs were used to seed the synthetic training generator. Two training configurations were evaluated: (i) a model trained exclusively on healthy (pre-injury) scans, and (ii) a model trained on the full longitudinal dataset including both pre- and post-injury time points. The remaining four CCI pigs were held out entirely for within dataset testing, ensuring subject-level separation between training and evaluation. To further assess generalizability, both trained models were applied without retraining to the second independent rotational cohort, representing a distinct injury mechanism and acquisition distribution. Performance was evaluated separately for healthy and injured cases across both cohorts, and the corresponding results are summarized in Figure 7.

Figure 7.

Comparison of Dice performance between models trained only on pre-injury data (dark bars) and models trained on both pre- and post-injury data (light bars). Results are shown separately for Rotational and CCI datasets under both pre-injury (left) and post-injury (right) test conditions. In all cases, the Combined model achieved significantly higher Dice scores than the Pre-Injury model (*p < 0.05, **p < 0.01, ***p < 0.001), demonstrating that including injury data during training improves generalization across both healthy and injured animals.

3.3. Model architecture for brain segmentation

The PIGSKIN framework employs a single-stage volumetric 3D U-Net trained and tested at native resolution. Each scan is cropped or padded to a standardized cube of 192³ voxels to ensure full head inclusion while preserving the original voxel size. This one-step design maintains anatomical fidelity and enables robust extraction of challenging regions such as the olfactory bulbs and ventral brain, which are often under-segmented by conventional methods.

The U-Net was implemented using the Voxelmorph framework ¹⁸. The network comprised 7 encoding–decoding layers with skip connections, each resolution level containing two 3 × 3 × 3 convolutional blocks with ReLU activations. The final layer produced two output channels with a voxelwise softmax activation, yielding probabilities for brain and non-brain tissue.

Connected-component filtering was applied post hoc to remove spurious islands, ensuring anatomically plausible final masks. For multimodal training, T1- and T2-weighted volumes were co-registered and processed jointly, with significant improvements observed over single-modality training (see Results).

3.4. Synthetic training data using clustering and spatial transformations

We developed a generative pipeline that dynamically produces synthetic image–mask pairs during training. In each iteration, expert-derived binary brain masks are first spatially transformed, after which foreground and background regions are clustered independently using Gaussian mixture models (GMMs) prior to label-to-image synthesis.

At the core of this pipeline is GMM-based clustering, which models the distribution of voxel intensities as a weighted sum of Gaussian components. The full synthetic augmentation workflow is illustrated in Figure 3, which depicts the sequence of spatial transformations, GMM clustering, and label-to-image synthesis steps that together generate anatomically consistent training pairs. Throughout this process, cluster identities corresponding to the foreground (brain) region remain intact and are consistently mapped to a binary brain mask, ensuring that the final synthetic image and its corresponding binary label are spatially and semantically aligned. All intensity-domain augmentations (e.g., Gaussian noise, Gaussian blur) were applied exclusively to synthesized images following label-to-image translation. Discrete label maps were never subjected to intensity perturbations or boundary blurring.

Figure 3.

Synthetic data generation pipeline; Overview of the synthetic training data generation and segmentation pipeline. (A) Input T1-weighted MRI scan. (B) Human-provided foreground binary annotation. (C) Foreground extraction: top row shows the extracted foreground region; bottom row shows clustering of foreground tissues via Gaussian Mixture Modeling (GMM). (D) Background clustering: top row shows background regions; bottom row shows background GMM clusters. (E) Combined label map merging clustered foreground and background regions. (F) Spatial augmentations applied to the label maps. (G) labels to image function to generate gray scale images. (H) and (I) example of a synthetic image from different views.

Given the expert-annotated binary label maps $S_i(x,y,z)$ that delineate brain from extracranial tissue. At each iteration of the training process one such map is randomly selected and perturbed by a spatial transform $\varphi$, composed of an affine and a non-linear component:

$$\varphi ={\varphi }_{\text{aff}}\circ {\varphi }_{\text{nonlin}}$$ (1)

The affine transformation ${\varphi }_{\text{aff}}$ includes random rotations (${\theta }_{\mathrm{x}}$, ${\theta }_{\mathrm{y}}$, ${\theta }_{\mathrm{z}}$) scalings ($s_x$, $s_y$, $s_z$) and translations ($t_x$, $t_y$, $t_z$):

$${\theta }_x,{\theta }_y,{\theta }_z\sim 𝒰(a_{\text{rot}},b_{\text{rot}})$$ (2)

$$s_x,s_y,s_z\sim 𝒰(a_{\text{sc}},b_{\text{sc}})$$ (3)

$$t_x,t_y,t_z\sim 𝒰(a_{\text{tr}},b_{\text{tr}})$$ (4)

The non-linear component ${\varphi }_{\text{nonlin}}$ is a diffeomorphic warp parameterized by a stationary velocity field (SVF). We first sample a coarse Gaussian random field ${\text{SVF}}^{\prime }\sim 𝒩(0,{\sigma }_{\text{SVF}})$, upsample to the target resolution ${\mathrm{r}}_{\text{HR}}$, and integrate with a scale-and-square approach:

$${\sigma }_{\text{SVF}}\sim 𝒰(0,b_{\text{nonlin}}),$$ (5)

$$SVF=\text{Resample}(SV{F}^{\prime };r_{HR}),$$ (6)

$${\varphi }_{\text{nonlin}}=\text{Integrate}(SVF)$$ (7)

The augmented label map is then given by:

$$L=S_i\circ \varphi .$$ (8)

3.5. Gaussian mixture modeling for clustering

Given the deformed label map $L$, intracranial and extracranial regions are clustered separately using Gaussian Mixture Models (GMMs). The labels produced by the GMM are not intended to correspond to anatomical structures. Instead, they function as local intensity strata, introduced purely to parameterize the label-to-image synthesis process. In this framework, anatomical meaning is deliberately deferred: biological validity is enforced only at the binary brain–non-brain level. Crucially, this constraint is maintained across all transformations, ensuring that the core anatomical boundary remains intact even as internal intensity structure is flexibly modeled. For a single-channel intensity $\mathrm{x}\in \mathrm{R}$, the probability distribution is:

$$p(x)=\sum _{k=1}^K{\pi }_k𝒩(x\mid {\mu }_k,{\sigma }_k^2)$$ (9)

where ${\pi }_k$ are mixture weights ${\sum }_{k=1}^K{\pi }_k=1$ and $N(x;{\mu }_k,{\sigma }_k^2)$ is a Gaussian density with mean ${\mu }_k$ and variance ${\sigma }_k^2$.In order to better capture complementary contrasts between T1- and T2-weighted images, we extended this framework to a two-dimensional GMM. For a voxel with intensity vector $\mathit{x}={(x_{\mathrm{T}1},x_{\mathrm{T}2})}^{\mathsf{T}}\in R^{𝟚}$ the joint distribution is modeled as:

$$p(\mathbf{x})=\sum _{k=1}^K{\pi }_k𝒩(\mathbf{x}\mid {\mu }_{\mathit{k}},{\Sigma }_{\mathit{k}}),\mathbf{x}=\left[\begin{array}{c}\hfill x_{T1}\hfill \\ \hfill x_{T2}\hfill \end{array}\right]$$ (10)

where each Gaussian component is parameterized by mean vector${\mu }_k\in R^{𝟚}$, covariance matrix ${\Sigma }_{\mathrm{k}}\in {\mathrm{R}}^{𝟚\times 𝟚}$, and mixture weight ${\pi }_{\mathrm{k}}$. Parameters were estimated with the expectation–maximization algorithm applied to expert-annotated seed data.

The clustered foreground and background components were subsequently passed to a spatial augmentation module to generate anatomically consistent variability, followed by label-to-image synthesis as described in the subsequent sections.

A label-to-image synthesis module then converted the perturbed label maps into realistic MRI volumes paired with binary masks. Because this procedure was executed on-the-fly during training, the network was continually exposed to novel examples, ensuring robustness and preventing overfitting, while providing effectively unlimited training data.

3.6. Spatial transformation on label maps

To improve robustness against anatomical variability and acquisition differences, synthetic label maps were perturbed with random affine and non-linear transformations before conversion to images. Transformations are explicitly constrained to avoid topological violations, including brain fragmentation, extrusion beyond the skull boundary, and inversion of foreground–background relationships. Affine perturbations included rotations uniformly sampled from ±180°, translations up to 6 mm, and anisotropic scaling up to ±10%. Non-linear deformations were simulated by applying smooth B-spline warps with a maximum displacement of 0.2 (in normalized units). Table 2 summarizes the ranges of all augmentation parameters. Notably, noise, blur, and cropping are not applied to the label maps; they are listed here for completeness and are used only later, during the label-to-image synthesis stage.

Table 2.

Parameters for synthetic data augmentation. For the combined model, spatial transformations included random rotations (±180°), shifts (≤6 mm), scaling (±0.1), and warping (0.2), along with blur (0.1), additive Gaussian noise (0.0–0.2), and cropping (probability = 1.0). Although full 180° rotations are unlikely in practice, this augmentation was included to promote rotational invariance and prevent orientation bias, similar to prior synthetic segmentation frameworks. In practice, the most effective rotations were within ±30–45°. Foreground and background components were perturbed with lower-intensity deformations to preserve tissue boundaries while enhancing variability.

	Label map transformation				Image transformation
Model	Rotation (°)	Shift (mm)	Scale	Warp	Blur	Noise	Crop Prob
Combined	180	6	0.1	0.2	0.1	0.0–0.2	1.0
Foreground	0	0	0.0	0.1	0.0	0.0	0.0
Background	0	0	0.0	0.05	0.0	0.0	0.0

3.7. Intensity augmentations of images derived from label maps

To simulate scanner-specific intensity non-uniformities and noise, voxel intensities were perturbed by additive Gaussian noise and multiplicative bias fields. These effects approximate coil inhomogeneity, B₁ bias, and random fluctuations in scanner gain that can vary across sites or acquisitions:

$$n\sim 𝒩(0,{\sigma }^2),\sigma \in [0,0.2],$$ (11)

where bias fields were generated by sampling a low-resolution Gaussian random field $b$, upsampling to full image size, and exponentiating to ensure positivity. The biased image was then computed as:

$$I_b(x)=I(x)\cdot \text{exp}(b(x)),$$ (12)

which induces smooth intensity inhomogeneities consistent with RF coil non-uniformity. To further randomize contrast, voxel intensities were rescaled to [0,1] and perturbed with a gamma transform,

$$I_g(x)=I_b{(x)}^{\gamma },\gamma \sim 𝒩(1,{\sigma }_{\gamma }^2),$$ (13)

allowing skew in the global histogram while preserving tissue separation ²⁰.

Resolution variability:

To make the segmentation robust to heterogeneous acquisition protocols, synthetic images were intentionally degraded to simulate variability in slice spacing and thickness. This step mimics real-world differences across scanners and studies, where porcine MRI is often acquired at variable resolutions and anisotropic voxels. After randomly selecting an imaging axis (axial, coronal, sagittal), the synthetic volume was blurred with a Gaussian kernel to emulate slice thickness, then downsampled by trilinear interpolation to a randomly chosen spacing:

$$I_{lr}={𝒟}_{𝓈}\left(G_{\sigma }\ast I_g\right),$$ (14)

where $G_{\sigma }$ denotes the slice-thickness kernel and ${𝒟}_{𝓈}$ is downsampling at spacing sss. The degraded image was finally upsampled back to 1 mm isotropic resolution, ensuring that the segmentation network learned to produce high-resolution predictions even from low-resolution inputs.

3.8. Model Training

Model training was seeded with a curated set of 19 manually annotated T1- and T2-weighted volumes from 7 unique pigs, including both healthy and injured conditions across multiple time points. Expert annotations were used to delineate brain foreground from extracranial background, which then served as input to our generative pipeline. Within this pipeline, tissue regions were decomposed into multimodal Gaussian mixture clusters, recombined into label maps, and perturbed through spatial and intensity transformations (including rotation, scaling, warping, noise, and blur). A labels-to-image synthesis module produced paired synthetic MR images and binary masks on the fly, producing effectively unlimited exemplars for training.

Training employed the Adam optimizer with an initial learning rate of 1 × 10⁻⁵, reduced by a factor of 0.95 after 20 epochs without improvement (minimum 1 × 10⁻⁷). Batch size was set to 1 due to GPU memory constraints, with 100 steps per epoch. The loss function was the soft Dice coefficient:

$${ℒ}_{\text{Dc}}=1-\frac{2{\sum }_i{p}_i{g}_i}{{\sum }_i{p}_i+{\sum }_i{g}_i+ϵ},$$ (15)

where ${\mathrm{p}}_{\mathrm{i}}$ and ${\mathrm{g}}_{\mathrm{i}}$ denote predicted and ground-truth voxel labels, respectively, and $\epsilon$ prevents division by zero. Each model was trained for unlimited number of epochs with early stopping at a convergence threshold of 10⁻⁴ Dice improvement.

Both U-Nets were trained from scratch using synthetic data. Synthetic samples do not preserve voxel-wise correspondence to any original scan and are regenerated stochastically at each training iteration. Soft Dice loss was used to optimize segmentation overlap. Training was validated on held-out synthetic samples with early stopping to prevent overfitting. The pipeline was implemented in TensorFlow and trained on a single NVIDIA A100 GPU.

3.9. Evaluation metrics for segmentation performance

Segmentation accuracy was quantified using the Dice similarity:

$$\text{Dice}(X,Y)=\frac{2\mid X\cap Y\mid }{\mid X\mid +\mid Y\mid }$$ (16)

where $\mathrm{X}$ is the predicted brain mask and Y is the expert annotation. Scores were reported as mean ± standard deviation across test subjects. Models were compared under different training/testing resolutions and modalities. Statistical significance between models was assessed using paired two-sided t-tests, with p < 0.05 considered significant.

For evaluation, model predictions were quantitatively compared against expert manual annotations to assess segmentation accuracy. Both experts independently delineated the brain masks, and their annotations served as ground truth references. PIGSKIN’s performance was benchmarked against these expert annotations as well as against several0020existing skull-stripping algorithms, enabling direct comparison across conventional and synthetic-based methods. To benchmark performance, PIGSKIN was evaluated alongside existing skull-stripping methods including BET, HD-BET, SynthStrip, and UniverSeg, using identical preprocessing and evaluation protocols. These baselines were selected to represent the current spectrum of automated brain extraction tools. BET (and its deep-learning variant HD-BET) are classical and widely adopted pipelines optimized for human MRI, providing a conventional reference. SynthStrip represents the newest state-of-the-art model trained on extensive synthetic human data, testing cross-species generalization. UniverSeg offers a universal few-shot segmentation framework that can adapt to unseen domains, making it an appropriate benchmark for assessing transferability. Finally, Original+ serves as an internal control trained directly on the limited real pig data with standard augmentations, isolating the contribution of our synthetic data generator. Together, these methods span the range from traditional to universal and synthetic-based algorithms, providing a fair and comprehensive comparison for PIGSKIN. In addition to whole-brain Dice scores, region-specific analyses were performed focusing on the olfactory bulbs and ventral brain, areas that are anatomically complex and often fail in conventional skull-stripping pipelines. These regions were defined a priori to assess model robustness in anatomically challenging zones, and qualitative and quantitative comparisons are reported in Section 4.

3.10. Test-time inference and resolution effects

All test scans were obtained from pigs that were not used to seed the synthetic training generator, ensuring subject-level separation between training and evaluation. At test time, we evaluated two inference strategies. First, following the SynthSeg pipeline ²⁰, input volumes were resampled to 1 mm isotropic resolution and normalized via min–max scaling using the 1st and 99th percentiles of intensities. The images were then cropped into a standardized 192³ voxel cube, and the U-Net produced the segmentation on the resampled volume.

Second, to evaluate the effect of downsampling-induced resolution loss — which occurs when images are interpolated from their native acquisition space to a coarser 1 mm grid — inference was performed directly on the native-resolution data (0.54 mm isotropic for the CCI cohort and 0.70 mm isotropic for the rotational-acceleration cohort). This one-step design preserved fine-scale anatomical details, such as the olfactory bulbs and ventral brain regions, which were otherwise prone to under-segmentation when evaluated exclusively at 1 mm resolution. Unless otherwise specified, all results reported in this study were obtained using native-resolution inference. Quantitative segmentation performance across training and testing resolutions is summarized in Table 3.

Table 3.

Impact of resampling on PIGSKIN performance (Dice ± SD).

TRAINING	TESTING	DICE SCORE (MEAN ± SD)
NATIVE (0.54 MM)	Native	0.968 ± 0.007
NATIVE	1 mm	0.962 ± 0.006
1 MM	1 mm	0.916 ± 0.011

For both inference modes, softmax probability maps were thresholded, and the largest connected component was retained to suppress spurious islands. The average runtime was approximately 12 s per scan for 1 mm inference and 18 s per scan for native-resolution inference, including resampling, network evaluation, and post-processing.

4. Results

4.1. Benchmarking against existing methods

Across all test sets, PIGSKIN achieved comparable to expert consistency accuracy, with Dice scores approaching those of human annotators when evaluated against Expert 1 (≈0.97; Figure 4). Although some disagreement persists with Expert 2 at anatomically ambiguous regions—particularly around the olfactory bulbs, which form bushy, continuous structures—overall inter-rater consistency remains high. Accordingly, inter-rater Dice largely reflects localized boundary uncertainty rather than systematic disagreement. Trained on spatially consistent synthetic variations, the model produces smoother and more internally consistent boundaries, which in turn leads to higher Dice scores relative to either expert when considered individually. Performance remained stable across cohorts and MRI modalities, demonstrating strong generalization. A U-Net trained solely on the original, non-synthetic images, including both healthy and injured brains with standard augmentations, failed to reach comparable accuracy, underscoring the importance of our synthetic data generator. Importantly, PIGSKIN’s performance approached the reference level of inter-rater variability (Expert 1 vs. Expert 2: 0.958 ± 0.006), effectively matching comparable to expert consistency accuracy (Fig. 4, Table 1). Notably, while Expert 2 annotations exhibited greater variability, PIGSKIN maintained stable predictions, suggesting that its apparent “errors” often reflect ambiguities in manual ground truth rather than algorithmic limitations. Together, these results establish PIGSKIN as the first porcine brain extraction framework to achieve human-expert performance, clearly surpassing existing automated methods. When the same model was trained and tested on 1 mm resampled images, accuracy dropped sharply (Dice = 0.73 ± 0.18). Visual inspection revealed systematic false negatives in anatomically complex regions. In several cases, one olfactory bulb was entirely missing from the predicted mask, highlighting the loss of critical detail when inference is performed at lower resolution.

Figure 4.

Comparison of segmentation performance across models using Dice score. Dice values are computed independently with respect to each expert annotation, with solid bars denoting evaluation against Expert 1 and hatched bars denoting evaluation against Expert 2. The proposed GMM-based PIGSKIN model demonstrates substantially improved performance relative to baseline methods (UniverSeg, BET, and Original+), achieving expert-level agreement with both annotators. The “E1/E2” bar represents inter-rater agreement between Expert 1 and Expert 2 and serves as a reference measure of annotation variability. Original+ refers to training on the original images with additional random augmentation (rotation and noise). Error bars indicate ±SD.

When trained and tested at native resolution, the one-step model achieved the highest Dice accuracy (0.968 ± 0.007). Evaluating the same network on 1 mm resampled images slightly reduced performance (Dice = 0.962 ± 0.006), with visible loss of fine structures such as the olfactory bulbs. Training and testing entirely on 1 mm resampled inputs further degraded performance (Dice = 0.916 ± 0.011).

PIGSKIN demonstrated highly consistent performance across T1- and T2-weighted test sets (Dice ≈ 0.968–0.970, Figure 5), indicating strong generalization across MRI contrasts (Figure 6). When trained jointly on T1 and T2 volumes, the model exhibited slightly improved stability, supporting a modality-agnostic framework that performs reliably irrespective of input sequence. This capability is particularly valuable in preclinical neuroimaging, where acquisition parameters vary across scanners, sites, and injury paradigms. Unlike existing pipelines constrained to a single contrast, PIGSKIN leverages synthetic multimodal clustering during training to achieve a unified segmentation solution that robustly accommodates both T1- and T2-weighted inputs.

Figure 5.

Example test case showing PIGSKIN segmentation results on T1- and T2-weighted MRI from the same pig. Red overlays represent the segmented region obtained using PIGSKIN, demonstrating consistent performance across both modalities.

Figure 6.

Performance of GMM-based models trained on T1-only, T2-only, or combined T1+T2 data, evaluated on separate T1 and T2 test sets. All models achieve high Dice scores across both modalities, with slightly improved consistency when trained with both T1 and T2. This demonstrates the generalization capability of the GMM framework across MRI contrasts.

4.2. Effect of injury inclusion on model generalization

Models trained on both pre- and post-injury data significantly outperformed those trained only on pre-injury cases. Dice scores improved from 0.91 to 0.96 in rotational datasets and from 0.95 to 0.97 in CCI datasets when tested post-injury. Exposure to pathological variability enhanced generalization across both healthy and injured animals (Figure 7).

PIGSKIN reliably segmented anatomically complex regions such as the olfactory bulbs and ventral brain, even under injury-induced disruption. Accurate separation was achieved in both healthy and injured conditions, outperforming conventional methods that often failed in these regions (Fig. 8).

Figure 8.

Examples of segmentation results across different injury models and datasets. The first row shows pre-CCI injury with clear separation of olfactory bulbs from ventral brain tissue (Axial view, box 1). The second row highlights post-CCI injury cases, where focal lesions are visible in the coronal (box 2) and sagittal (box 3) views. The third and fourth rows display pre- and post-diffuse injury scans, respectively, with box 4 and 5 illustrating the olfactory region where boundaries are irregular and complex.

Overall, qualitative inspection confirmed accurate delineation of the brain boundary, including challenging regions such as the olfactory bulbs and ventral surface (Figure 5). PIGSKIN preserved fine-scale anatomical detail without including extracranial tissues.

5. Discussion

This work introduces PIGSKIN, a synthetic data–driven framework for porcine brain extraction. PIGSKIN achieved Dice scores approaching inter-rater reliability, substantially outperforming conventional human brain segmentation pipelines such as BET and UniverSeg^9,21, which are designed and trained exclusively on human MRI and fail when applied to porcine anatomy. These tools typically assume smooth cortical boundaries, limited extracranial tissue, and consistent head shape, assumptions violated in pigs due to thick subcutaneous fat, elongated snout, and ventral brain extension into the olfactory and cerebellar regions.

The term comparable to expert consistency refers to parity with the agreement observed between two trained raters (Dice ≈ 0.958 ± 0.006). PIGSKIN’s performance (Dice ≈ 0.97) thus exceeds the variability inherent to human annotation, indicating that the model is as consistent as, or more consistent than, manual delineation. Most discrepancies between experts arose near the ventral brainstem and olfactory bulbs, regions where boundaries are diffuse or anatomically irregular. PIGSKIN preserved these structures reliably, reducing both false negatives and inter-rater disagreement (Figure 4).

Accurate segmentation of the olfactory bulbs and ventral brain has historically been difficult because these regions exhibit irregular geometry, contact extracranial fat, and vary in intensity due to air–tissue interfaces. Conventional human skull-stripping pipelines often truncate or merge these areas because their priors enforce compact, convex masks. In contrast, PIGSKIN’s synthetic multimodal augmentation explicitly exposed the network to such anatomical extremes, elongated bulbs, ventral flattening, and heterogeneous soft-tissue boundaries, allowing it to learn invariant spatial cues. As shown in Figure 8, the model maintained consistent delineation across both healthy and injured animals, outperforming all conventional methods in these anatomically complex regions.

Human pipelines failed largely because their priors enforce compact, convex brain masks and expect minimal extracranial intensity overlap. In pigs, extracranial fat, muscle, and air pockets near the sinuses break these assumptions, causing under-segmentation or leakage into soft tissue. PIGSKIN overcomes this by decoupling intracranial and extracranial intensity modeling during synthetic data generation. Foreground and background are clustered independently and re-combined under randomized spatial and intensity perturbations, generating anatomically plausible yet diverse examples (Figure 3). This augmentation strategy exposes the model to the full spectrum of porcine variability, preventing overfitting to a single anatomy and enabling robust generalization across scanners and injury paradigms.

While prior work often downsampled to standardized resolutions for convenience, porcine brains contain small high-contrast structures easily degraded by interpolation. We found that maintaining native spatial resolution preserved anatomical fidelity rather than merely improving scores. The olfactory bulbs and ventral brain, key sites of deformation in TBI, are particularly sensitive to partial-volume effects. Training at native resolution thus ensures quantitative accuracy for downstream volumetric and morphometric analyses while still generalizing well to coarser datasets (Table S1).

Together, these design choices, synthetic multimodal augmentation, independent tissue modeling, and preservation of native spatial detail, enabled a model that generalizes across breeds (Yorkshire, Landrace), scanners, and injury types without large annotated datasets. Beyond pigs, this framework outlines a strategy for data-limited domains: learn anatomical variability through controlled synthesis rather than manual labeling.

In sum, PIGSKIN achieves reproducible, comparable to expert consistency brain extraction with minimal supervision and provides a blueprint for anatomically grounded synthetic training in preclinical and translational neuroimaging.

Limitations and Future Work:

Several limitations should be acknowledged. The evaluation cohort was modest compared with human benchmarks, which may limit generalizability and prevent stratification by injury severity or site. Synthetic augmentation, while powerful, was seeded from up to seven pigs, potentially constraining anatomical diversity. Training data were largely derived from CCI animals, raising the possibility of model bias toward this injury model ^22,23. Broader validation across breeds, injury types, and sites will be essential.

Methodologically, augmentation was largely slice-based, which may limit 3D spatial continuity. Future work should incorporate volumetric generative models, hybrid CNN–transformer architectures, and self-supervised pretraining. Finally, expert annotations are subject to variability, especially in injured regions. Consensus- or probabilistic-based ground truth would strengthen evaluation frameworks.

6. Conclusion

PIGSKIN delivers the first automated, comparable to expert consistency solution for porcine brain extraction. By uniting synthetic data generation with a two-stage deep learning framework, it achieves robust performance across modalities, injury models, and datasets. Beyond filling a critical gap in preclinical neuroimaging, PIGSKIN illustrates how synthetic training can enable scalable and reproducible pipelines in resource-limited domains. This tool lays the groundwork for quantitative porcine TBI research and strengthens the translational bridge to human neuroimaging. To accelerate adoption, we have released the full code and trained models publicly at https://github.com/dadashkarimi/PIGSKIN.

Figure 2.

Overview of the U-Net architecture for image segmentation. The contracting path (left) applies convolution and max pooling to capture context, while the expansive path (right) uses upsampling and skip connections to recover spatial resolution. Skip connections concatenate feature maps from corresponding layers, enabling precise localization in the final segmentation output.

Data Availability Statement

The trained model weights and source code for the PIGSKIN framework are publicly available at https://github.com/dadashkarimi/PIGSKIN. The imaging data used in this study are not publicly available due to ethical and institutional restrictions related to animal research but may be made available from the corresponding author upon reasonable request and subject to institutional approval.