Mitochondria Segmentation

Task Type

Segmentation

Overview

The Mitochondria Segmentation model performs pixel-wise classification of mitochondria in electron microscopy (EM) images.
It supports both 2D slice-wise model (trained on CEM-Mitolab dataset) and 3D volumetric-input model (trained on MitoEM dataset).

This task outputs a labeled mask with the following classes:

  • Background
  • Mitochondria

Available Solutions

1. ViT-L-2D

Configuration

- name: "ViT-L-2D"
  note: "trained on CEM-MitoLab, no 3D fusion"
  datatypes:
    - 2
    - 3
  img_z: 1
  backbone_config: null
  weights: "weights/mito-seg/905953_model_199.pt"

Characteristics

  • img_z = 1 (2D model)

  • Accepts:

  • 2D input

  • 3D input (processed slice-wise)
  • No inter-slice fusion
  • Suitable for fast inference and datasets without strong Z-continuity
  • 2D EM slice segmentation
  • Large-scale screening
  • When Z-alignment consistency is uncertain

2. ViT-L-3D

Configuration

- name: "ViT-L-3D"
  note: "trained on MitoEM, with 3D fusion, available when z_dim >= 16"
  datatypes:
    - 3
  img_z: 16
  backbone_config: null
  weights: "weights/mito-seg/909906_model_239.pt"

Characteristics

  • img_z = 16 (3D model)
  • Accepts 3D input only
  • Requires:
  • Z-dimension ≥ 16

Output

The model produces a segmentation mask with:

Label ID Class Name
0 Background
1 Mitochondria

Model Selection Guidance

Scenario Recommended Solution
Single 2D slice analysis ViT-L-2D
Thin Z-stack (< 16 slices) ViT-L-2D
Thick volumetric dataset (≥ 16 Z) ViT-L-3D

Choose the 3D model when volumetric coherence is critical and sufficient Z-depth is available. Use the 2D model for flexibility, speed, or limited Z-resolution datasets.

Upcoming Updates

We are actively improving the mitochondria segmentation models. Planned updates include:

  1. Training on MitoEM 2.0 (3D Model)
    A new 3D segmentation model will be trained on the MitoEM 2.0 dataset to enhance volumetric accuracy, robustness, and generalization across diverse EM imaging conditions.

  2. Post-processing for Cross-Dimensional Consistency
    We will introduce post-processing techniques to improve consistency between:

  3. Slice-wise 2D segmentation outputs
  4. Volumetric 3D segmentation outputs

These enhancements aim to reduce inter-slice discontinuities and improve structural coherence in reconstructed 3D volumes.

These updates will be released in future versions of Napari-OmniEM.