Few-shot segmentation and fiber tractography of human vagus nerve using 3D-MUSE imaging

Abstract

We are dissecting and imaging 100 human cadaver nerves with unprecedented range of anatomical coverage (from brainstem to abdomen) and imaging modalities. Our teams used 3D serial block-face Microscopy with Ultraviolet Surface Excitation (3D-MUSE) to image, visualize, and quantify the morphology and microanatomy of the human vagus nerve, providing three-dimensional insights into its structure and functional organization. We prepared 3-mm and 5-mm-long samples of human cervical vagus and median nerve using various staining and embedding techniques before imaging with 0.9-μm in-plane resolution and between 3-μm and 12-μm slice thickness. Staining quality varied across samples thus requiring training of a sample-based neural network rather than using a generalized analysis algorithm. We used few-shot learning to segment the fascicles, perineurium, and epineurium regions. We trained a 2D U-Net architecture with 4–8% of each sample’s images before applying to a held-out test set. Performance achieved a mean Dice score range of 0.85±0.10 and 0.93±0.05 across various 3D-MUSE samples. We also investigated an initial pre-training step of the U-Net model to improve segmentation performance. Pre-training enabled the segmentation model to have better awareness of splitting fascicles in the held-out test set. From sample segmentation predictions, morphologic metrics such as nerve diameter, fascicle count, fascicle area, fascicle diameter, and perineurium thickness are calculated. Nerve fiber tractography from sample images highlight dynamic fascicle organization throughout 3-mm nerve samples These results demonstrate the importance and success of sample-based training for segmentation and nerve fiber tractography, with further training anticipated to yield even better outcomes.

Summary:

We use 3D serial block-face Microscopy with Ultraviolet Surface Excitation (3D-MUSE) to image and analyze the human vagus nerve, contributing to the first vagal connectome. This technique bridges the gap between microCT and histology, providing detailed visualization of small fascicle merging/splitting and 3D perineurium boundaries. We developed a few-shot deep learning model for segmenting anatomical structures, using samples with distinct staining protocols. The 2D U-Net model trained with 4–8% manually annotated sample images achieved Dice scores greater than 0.85 and accurate morphometric analysis, demonstrating the potential for automatic segmentation and nerve fiber tractography with further training.

1. INTRODUCTION

Peripheral neuromodulation, which involves external stimulation to modulate nerve activity, is an emerging field in medical research with applications in treating conditions like epilepsy¹, depression², and obesity³. However, its therapeutic impact remains relatively modest. With vagus nerve stimulation (VNS), only 50% of patients demonstrate substantial improvement, largely due to challenges in targeting and associated adverse effects^4–6. To enhance these therapies, computational modeling can optimize stimulation parameters and cuff designs, but accurate anatomical data on nerve morphology, such as nerve fiber organization and perineurium thickness, is essential^7–11. Our team is conducting a detailed dissection and imaging of 50 human cadaver nerves, covering an extensive anatomical range from the brainstem to the abdomen using a variety of advanced imaging modalities. Previously, we developed three-dimensional Microscopy with Ultraviolet Surface Excitation (3D-MUSE) to image detailed visualization of nerve structures via serial block-face imaging¹². 3D-MUSE provides complementary information compared to histology and microCT: both 3D-MUSE and histology can resolve small fascicles and perineurium, but 3D-MUSE can do so in three dimensions rather than only in cross section, whereas microCT provides three-dimensional morphology, but it cannot reliably detect perineurium boundaries and its lower resolution fails to capture small fascicles. Further, tractography with 3D-MUSE data provides insights into functional organization of the neurons in three dimensions, for which microCT does not have contrast or resolution. To enable the latter analysis, our group also developed NerveTracker¹³, a Python-based software that tracks nerve fiber groups within 3D-MUSE images using structure tensor¹⁴ and optic flow¹⁵ algorithms.

While 3D-MUSE presents many advantages other modalities find challenging, a couple challenges arise in terms of image analysis. For example, to calculate morphometrics such as perineurium thickness and fascicle diameter for computational modeling parameters, segmentation of the perineurium and fascicle tissues is required from 3D-MUSE samples. However, since 3D-MUSE imaging can generate between 40–200 GB of data per sample, efficient automatic segmentation is necessary. Additionally, staining variation and quality across samples, anatomical region, and cadaver body requires sample-based deep learning model training rather than a generalized global neural network. Thus, in this study we explored a sample-specific few-shot learning technique using less than 10% of labeled sample images for multiclass segmentation training and testing of the fascicles, perineurium, and epineurium from human vagus nerve samples. Additionally, we performed fiber tractography on the same human vagus nerve samples to analyze their dynamic fascicular organization. These efforts indicate the potential of sample-based training for segmentation and nerve fiber tractography.

2. METHODS

Dataset

This study prepared, imaged, and analyzed three human cervical vagus nerve samples and one human median nerve sample from different human cadavers and staining protocols. Sample 1, a mid-cervical nerve sample, was stained with Ehrlich’s hematoxylin and rhodamine, while Sample 2, a left cervical nerve sample, received an initial phosphotungstic acid stain before Sample 1’s staining. Sample 3, a mid-cervical nerve sample was stained with luxol fast blue (LFB) and rhodamine B before fixation in dimethyl sulfoxide for 48 hours. Finally, Sample 4, a median nerve sample was stained with methylene blue, rhodamine B, and LFB. We collected 1000 images from Sample 1 of size [3000, 4000] in dimension with 3-μm slice spacing. Sample 2 images were [6300, 8400] in dimension with 3-μm slice spacing for a total of 1000 images as well. From Sample 3, 435 images were collected of size [6300, 8400] with 12-μm slice spacing, and 352 images of size [3900, 6000] were collected from Sample 4 with 9-μm slice spacing. Table 1 below reports these characteristics as well as the number of images used for training and testing during segmentation experiments. Ground truth multiclass labels were selected and annotated at evenly spaced images, between 30–90μm apart depending on the sample. To generate multiclass labels, the fascicles, perineurium, and epineurium were manually annotated within each selected image. All experiments incorporated the MONAI (Medical Open Network for AI) Python framework

Table 1.

Information on video and audio files that can accompany a manuscript submission.

Sample	Total Images	Training Images	Testing Images	Image Spacing
1	1000	40	10	3-μm
2	1000	80	20	3-μm
3	250	20	5	12-μm
4	352	30	5	9-μm

Network Architectures

To conduct sample-specific few-shot learning, we trained a U-Net architecture with residual nets¹⁶ for supervised semantic segmentation. U-Net’s skip connections allow the network to recover full spatial resolution in its decoding layers, while the addition of residual nets help to overcome vanishing gradients and improve feature extraction¹⁷. A unique U-Net model was trained for and by each sample. We also experimented with an initial pre-training step to improve segmentation performance and take advantage of the vast quantity of imaging data from 3D-MUSE samples. Using all unlabeled images from a given sample, 20 patches of [75, 75] were dropped out from each image and used to pre-train U-Net by comparing the model’s reconstruction of the dropped-out patched with the original images via and L1 loss function. The pre-trained model’s weights initialized the U-Net architecture before fine-tuning using the few-shot learning pipeline described earlier. To generate segmentation predictions for an entire sample, A 2D U-Net model was trained on all labeled images from a given sample before the model was applied to every sample image. Once the three tissue classes (fascicle, perineurium, and epineurium) were segmented from the sample, we calculated nerve morphology metrics.

Performance Assessment

We assessed model performances on held-out test set images by calculating the Dice Coefficient and Hausdorff distance within each class (fascicle, perineurium, and epineurium) from a given multiclass segmentation prediction, and by taking an average of the three class Dice Coefficients and Hausdorff distances, respectively. Furthermore, we used segmentation predictions to calculate the nerve diameter, fascicle counts, fascicle area, fascicle diameter, and perineurium thickness.

Fiber Tractography

The NerveTracker software incorporates two flow estimation algorithms tailored for block-face microscopy datasets: optic flow tractography and structure tensor tractography. The optic flow method, based on the Lucas-Kanade algorithm¹⁸, tracks points across 3D-MUSE image stacks by treating them as video frames, estimating flow vectors at each pixel under the assumption of small displacements and constant brightness. This method involves a pyramidal approach for handling larger displacements fields between consecutive images by downsampling images iteratively. In contrast, the structure tensor analysis estimates a 3D orientation vector at each voxel by minimizing the sum of squared differences between a neighborhood’s displacement derived from image intensity variations within a local neighborhood. This approach uses Eigen decomposition to identify the optimal orientation vector, which is then scaled to determine the new voxel location across image planes. Both methods generate streamlines by connecting the old and new pixel locations, with sub-pixel accuracy achieved through interpolation. Streamline tracking is terminated if it diverges significantly from the nerve’s long axis, and streamlines are color-coded based on their originating regions of interest in the input mask.

3. RESULTS

All samples investigated varied in image quality, staining protocol, nerve location, and cadaver body. Thus, we trained sample-specific few-shot multiclass segmentation models – a unique model was developed using only a single sample’s images. Each model was then applied to the held-out test set selected from each corresponding sample. Figure 1 illustrates the unique sample images and their corresponding multiclass segmentation results of the sample-based deep learning models compared to the manually annotated labels. Table 2 compliments these results quantitatively by reporting the Dice Coefficient and Hausdorff distance. These metrics demonstrate that even with less than 10% of sample images used for training, our U-Net model can successfully segment the three tissue classes on images 30–120-μm away from a ground truth label. Figure 2 illustrates how pre-training helped provide better awareness of the splitting fascicles during the fine-tuning training and in the predicted segmentation as indicated by the yellow circles.

Figure 1.

Example images (a), multiclass predictions (b), and corresponding ground truth labels (c) from four human vagus (1–3) and median (4) nerves prepared using various staining, embedding, and imaging protocol. In row 1, epineurium (purple), perineurium (blue), and fascicles (red). In row 2–4, epineurium (red), perineurium (purple), and fascicles (blue).

Table 2.

Quantitative segmentation performance metrics for various 3D-MUSE samples

Sample	Images	Train	Test	Dice Coefficient	Hausdorff Distance
1	1000	40	10	0.93 ± 0.05	204 ± 185
2	1000	80	20	0.87 ± 0.10	83 ± 56
3	435	30	5	0.87 ± 0.08	203 ± 202
4	352	20	5	0.85 ± 0.10	354 ± 196

Figure 2.

Multiclass segmentation demonstrating the benefit of pre-training from two example images. Yellow circles indicate instances where splitting fascicles were recognized by an initially pre-trained model.

From sample segmentation predictions, nerve morphology metrics were calculated, including nerve diameter, fascicle count, fascicle area, fascicle diameter, and perineurium thickness. All metrics except for nerve diameter are calculated as an average for a given image’s segmentation. Figure 3 illustrates these metrics derived from a 3-mm mid-cervical vagus nerve. Two trends of note are that as the number of fascicles decrease, the perineurium thickness increases, and as the number of fascicles increase, the perineurium thickness decreases. Furthermore, Figure 4 represents a positive correlation between perineurium thickness and fascicle diameter throughout the same sample, a similarly reported trend by Grinberg et al¹⁹. These results indicate that as the number of fascicles increase, likely by splitting small-to-medium fascicles, the perineurium tissue which must encompass all fascicles, thins and decreases in average thickness overall.

Figure 3.

Perineurium thickness versus fascicle diameter along a 3-mm human mid-cervical vagus nerve

Figure 4.

Nerve morphology metrics calculated from multiclass segmentation of a 3-mm human mid-cervical vagus nerve imaged with 3D-MUSE. Metrics include nerve diameter, fascicle count, fascicle area, fascicle diameter, and perineurium thickness along the nerve sample length

Finally, after performing tractography using structure tensor algorithm for Sample 1 and optic flow algorithm for Sample 2, both mid-cervical vagus nerves we see the dynamic change of fascicular makeup throughout each sample’s ROI. Figure 5 demonstrates this – colored groups of streamlines that make up one fascicle at the 0-mm slice of the sample stack not only split and migrate to new locations but also redefine new fascicular organizations throughout the 3-mm stack. We can also observe that as new fascicular organizations form during splitting and merging events, streamlines are not inter-mixing but rather maintain their original organization.

Figure 5.

Nerve fiber tractography streamlines clipped at example 3D-MUSE images throughout Sample 1 and Sample 2 nerve sample.

4. CONCLUSIONS

This work demonstrates success of sample-based training for multiclass segmentation using varying human vagus nerve samples capturing different image quality, staining protocol, deep learning, and pre-training techniques. All sample-based training models performed accurately and comparably within their sample’s held-out test sets with Dice scores greater than 0.85. Few-shot learning benefited from pre-training with unlabeled images as this enhanced awareness of splitting fascicles. Morphology analysis demonstrated direct correlation trends between perineurium thickness and fascicle diameter, and inverse correlation trends between perineurium thickness and the number of fascicles. Lastly, structure tensor and optic flow tractography algorithms illustrate the complexity and dynamic behavior of fascicular makeup within a 3-mm long human vagus nerve sample. We intend to continue investigating the optimization of sample-based segmentation models by training on more and varying samples and using these segmentations to anatomically constrain and enhance the tractograms of corresponding samples.