Architecture¶

This page describes the modeling scope, target structures, and architectural design choices of MicroStructFormer.
The framework is explicitly designed to focus on subcellular and periaxonal ultrastructures within myelinated fibers, and its modeling choices reflect this targeted scope.

Target structures¶

MicroStructFormer models a set of ultrastructural compartments that are central to axon–myelin–organelle interactions but are not reliably captured by conventional axon–myelin segmentation pipelines.

In the current implementation, the primary target structures include:

Mitochondria inside axons
Mitochondria located within the axoplasmic compartment, associated with axonal metabolic demand and transport processes.
Periaxonal space
The narrow compartment between the axonal membrane and the inner myelin lamella, which plays a key role in axon–glia interactions and metabolic exchange.
Mitochondria outside axons
Mitochondria located in the extra-axonal space, including glial or interstitial regions, which are treated as a distinct semantic class to avoid conflation with axonal organelles.
Abnormal or non-compact structures
Morphologically heterogeneous regions associated with disrupted, non-compact, or pathological ultrastructure, including irregular myelin-associated features.
Mitochondria-like organelles
Membrane-bound or electron-dense organelles that exhibit partial morphological similarity to mitochondria but do not meet strict annotation criteria for canonical mitochondria.

These structures are modeled as semantic classes and are designed to be analyzed both at the image level and in association with individual axonal instances defined by external segmentation.

Model architecture¶

MicroStructFormer adopts transformer-based semantic segmentation architectures to model fine-grained ultrastructural semantics in electron microscopy images.
The design follows established principles for joint multi-class segmentation with global contextual reasoning, as exemplified by Mask2Former and related transformer-based frameworks.

By leveraging hierarchical feature representations and attention-based contextual integration, the model is well suited to capturing heterogeneous subcellular morphologies and spatially extended structures.
Architectural details largely follow existing formulations and are not re-derived here. See the Mask2Former paper (arXiv) and the official implementation (GitHub) for further infomation.

What is not modeled¶

To avoid ambiguity regarding the scope of the framework, it is important to explicitly state which structures and tasks are not modeled by MicroStructFormer in its current form.

MicroStructFormer does not perform:

Axon and Myelin segmentation
Axonal area and myelin wrapping are handled by external tools and are not explicitly learned or predicted within the MicroStructFormer framework.
Biophysically explicit modeling
The framework does not simulate myelin growth, axonal transport, or metabolic processes, and should not be interpreted as a mechanistic or dynamical model.
General-purpose EM segmentation
MicroStructFormer is not intended to replace comprehensive EM segmentation tools that aim to label all cellular components within a field of view.

MicroStructFormer is designed to integrate with existing axon–myelin segmentation tools while providing specialized modeling of subcellular and periaxonal ultrastructure that is difficult to capture with generic pipelines.

Notes on extensibility¶

The current set of target structures reflects the primary focus of the present study.
Additional compartments or modeling tasks may be incorporated in future versions, provided that their inclusion does not compromise the interpretability, reproducibility, or modular structure of the pipeline.

Model Availability¶

Download detail please refer to the Model Zoo section. If the provided MicroStructFormer models do not adequately capture the characteristics of your dataset or experimental setting, we encourage you to contact us to discuss potential model adaptation or training on custom data.