Source-linked AI summary
AxonCallosumEM Dataset: Axon Semantic Segmentation of Whole Corpus Callosum cross section from EM Images
Ao Cheng, Guoqiang Zhao, Lirong Wang, Ruobing Zhang
TL;DR
Whole-corpus-callosum EM datasets with dense, fully proofread labels for axons and myelin sheaths are lacking, despite complex morphologies that challenge existing methods. This paper presents the AxonCallosumEM dataset and adapts SAM for EM segmentation as EM-SAM. The authors report that EM-SAM achieves state-of-the-art performance on their benchmark.
Problem
Whole-corpus-callosum EM reconstructions with dense, fully proofread ground truth are lacking, while complex morphologies and special cases challenge existing segmentation methods.
Source: §1 · p. 2Method
The authors construct AxonCallosumEM with dense annotations of the entire mouse corpus-callosum cross-section and fine-tune SAM for automatic EM segmentation as EM-SAM.
Source: §2 · p. 3, §2.2 · p. 4, §3 · p. 5Results
EM-SAM achieves state-of-the-art performance on AxonCallosumEM and handles large objects, long-range morphology, and demyelination, although some segmentation errors persist.
Source: §4.2 · p. 6Takeaways & Limitations
The densely annotated dataset may support method evaluation, pre-training, cross-model comparisons, and biomedical analyses across mouse models.
Source: §5 · p. 6
Abstract
from arXiv · showhide
The electron microscope (EM) remains the predominant technique for elucidating intricate details of the animal nervous system at the nanometer scale. However, accurately reconstructing the complex morphology of axons and myelin sheaths poses a significant challenge. Furthermore, the absence of publicly available, large-scale EM datasets encompassing complete cross sections of the corpus callosum, with dense ground truth segmentation for axons and myelin sheaths, hinders the advancement and evaluation of holistic corpus callosum reconstructions. To surmount these obstacles, we introduce the AxonCallosumEM dataset, comprising a 1.83 times 5.76mm EM image captured from the corpus callosum of the Rett Syndrome (RTT) mouse model, which entail extensive axon bundles. We meticulously proofread over 600,000 patches at a resolution of 1024 times 1024, thus providing a comprehensive ground truth for myelinated axons and myelin sheaths. Additionally, we extensively annotated three distinct regions within the dataset for the purposes of training, testing, and validation. Utilizing this dataset, we develop a fine-tuning methodology that adapts Segment Anything Model (SAM) to EM images segmentation tasks, called EM-SAM, enabling outperforms other state-of-the-art methods. Furthermore, we present the evaluation results of EM-SAM as a baseline.
1 Introduction
Whole-corpus-callosum EM reconstruction lacks dense, fully proofread ground truth, while the region’s axon morphologies and distributions pose challenges for existing segmentation methods. AxonCallosumEM addresses this gap with a large-scale dataset and examines CNN-, transformer-, and natural-image-model approaches to segmentation.
1 Introduction
- The corpus callosum connects the cerebral hemispheres and contains extensive axon bundles with complex morphology, while public EM datasets have largely focused on gray matter.Only a few unreleased EM datasets focus on corpus-callosum white matter.Source: §1 · p. 2
- Existing whole-corpus-callosum EM reconstructions lack dense, fully proofread ground truth for myelinated axons and myelin sheaths, and existing methods struggle with the region’s varied morphologies and special cases.Challenges include scale differences, dense axons, demyelination, Nodes of Ranvier, and long-range axons with indistinct boundaries.Source: §1 · p. 2
- AxonCallosumEM is a large-scale 2D axon semantic-segmentation benchmark with over 600,000 patches and comprehensive ground truth for myelinated axons and myelin sheaths.The dataset also showcases the entire cross-section and varied morphologies, supporting comparative studies across mammalian species and mouse models.Source: §1 · pp. 2–3
1.1 Related work
- Convolutional methods include binary-map, affinity-map, contour, and distance-target approaches, but their limited perceptual fields hinder segmentation of large-scale cellular structures.Vision transformers use long-range dependencies to address such issues.Source: §1.1 · p. 3
2 AxonCallosumEM dataset
AxonCallosumEM covers the mouse corpus callosum’s entire cross-section with dense annotations of myelinated axons and myelin sheaths. Its analysis reports a mean g-ratio of approximately 0.7 and a distribution unlike known anatomy in other mouse models.
2 AxonCallosumEM dataset
- AxonCallosumEM encompasses the entire mouse corpus callosum cross-section with dense annotations of myelinated axons and myelin sheaths.The annotations are intended to support segmentation-method evaluation and biomedical analysis.Source: §2 · p. 3
2.1 Dataset discription
- The dataset uses Mecp2-mutant Rett Syndrome mice and contains over 600,000 patches of 1024 × 1024 pixels at 4 nm per pixel.The image covers the genu, body, and splenium and is arranged as 448 by 1408 patches.Source: §2.1 · p. 4
2.2 Dataset annotation
- Annotations began with manual labeling across the x-axis, followed by SAM-based prediction along the y-axis and proofreading with manual error correction.The team iteratively expanded labeled patches along the y-axis to reconstruct the full cross-section.Source: §2.2 · p. 4
2.3 Dataset analysis
- For aggregate analysis, labels were downsampled to 16 nm per pixel and masked to the corpus-callosum region; reported metrics include axon diameter, density, volume fractions, and g-ratio.The g-ratio is the ratio of axon diameter to myelinated-fiber diameter.Source: §2.3 · p. 4
- The RTT mouse model’s mean g-ratio was approximately 0.7, and its distribution differed from known anatomy in other mouse models.Figure 3 maps g-ratio alongside axon diameter, density, and axon and myelin volume fractions across the cross-section.Source: Figure 3 · p. 4, §2.3 · p. 5
3 Methods
The methods compare established EM segmentation approaches with vision-transformer models and adapt SAM for automatic EM segmentation. The comparison includes pretrained MAE and BEiT models and an EM-SAM variant without prompt-dependent components.
3 Methods
- The methods compare CNN-based EM approaches that predict binary maps, affinity maps, contours, or distance targets, and use a vision-transformer model to address CNN limitations.The cited challenges include scale differences, Nodes of Ranvier, and long-range axons.Source: §3 · p. 5
- The comparison includes publicly available pretrained MAE and BEiT vision transformers and a fine-tuned SAM image encoder adapted as EM-SAM.The SAM prompt encoder and image-encoder neck are removed because the automatic segmentation pipeline uses no prompts.Source: §3 · p. 5
- To address CNN-based methods’ challenges with scale differences, Node of Ranvier, and long-range axons, the authors employed a ViT-based model, compared MAE and BEiT, and fine-tuned SAM’s image encoder to EM-SAM.They removed SAM’s prompt encoder and image-encoder neck because the automatic segmentation pipeline does not involve prompts.Source: §3 · p. 5
4 Experiments
Experiments compare segmentation models on AxonCallosumEM, with EM-SAM achieving state-of-the-art performance while retaining some semantic errors. The evaluation also examines pretrained-model gains, model settings, and qualitative failure cases.
4.1 Implementation Details
- ViT models share a decoder, augmentation, and training schedule; EM-SAM uses 1024×1024 inputs, versus 224×224 for MAE and BEiT.Training uses SGD, BCE loss, 200,000 iterations, batch size 2, and two NVIDIA 3090 GPUs; mIoU is the evaluation metric.Source: §4.1 · p. 5
4.2 Benchmark results on AxonCallosumEM dataset
- Models are selected using a validation split and evaluated on a test split, with the genu, body, and splenium assigned to training, validation, and testing.The comparison includes UNet, fine-tuned MAE and BEiT, randomly initialized ViT, and EM-SAM.Source: §4.2 · p. 5
- ViT improves mIoU by 0.028 over UNet, with self-attention capturing long-range dependencies despite a smaller input size.UNet has the lowest mIoU and struggles with large or heterogeneous objects, though it segments common-sized objects accurately.Source: §4.2 · p. 5
- Figure 4 contrasts EM-SAM's long-range segmentation successes with CNN failures on large or complex objects, and shows residual EM-SAM and untuned-SAM errors.EM-SAM can misclassify myelin gaps or thick-membraned nuclei as axons; untuned SAM does not categorize all axons and myelin sheaths.Source: Figure 4 · p. 6
- EM-SAM achieves state-of-the-art performance versus MAE and BEiT, attributed to its much larger original training dataset and 4× larger input.It segments large objects and handles long-range morphology and demyelination, but still confuses some myelin gaps or thick-membraned nuclei with myelinated axons.Source: §4.2 · p. 6
- Pretrained MAE and BEiT improve mIoU by 0.019 and 0.015, respectively, over training from scratch.Some EM-SAM errors involve unclear boundaries and morphology.Source: §4.2 · p. 6
5 Conclusion
The conclusion presents AxonCallosumEM as a large-scale, densely annotated dataset that exposes current methods' limits on complex morphologies. It also describes possible dataset applications and EM-SAM's approach to adapting natural-image models for EM processing.
5 Conclusion
- AxonCallosumEM is a large-scale dataset for myelinated-axon and myelin-sheath segmentation that reveals current methods' limitations on complex morphologies.The conclusion presents EM-SAM as an approach to adapting large natural-image models for automatic, high-throughput EM image processing.Source: §5 · p. 6
- The dataset may support pre-training, evaluation of new methods, adaptation between natural and biomedical image modeling, and biomedical analyses across mouse models.Source: §5 · p. 6