Hierarchical Encoder-Decoder With Soft Label-Decomposition for Mitochondria Segmentation in EM Images.

Area of Science:

• Cell Biology
• Computational Biology
• Image Analysis

Background:

• Accurate semantic segmentation of mitochondria from electron microscopy (EM) images is crucial for obtaining reliable morphological statistics.
• Automated delineation of mitochondria with diverse shapes and complex backgrounds presents significant challenges in accuracy.

Purpose of the Study:

• To develop an advanced deep learning model for precise semantic segmentation of mitochondria in EM images.
• To address the challenges of irregular mitochondrial shapes and complex image backgrounds.

Main Methods:

• A hierarchical encoder-decoder network (HED-Net) with a three-level nested U-shape architecture was developed to capture contextual information.
• A novel soft label-decomposition strategy was introduced, decomposing manual labels into subcategory-aware maps based on mitochondrial ovality.
• The HED-Net was trained using the original label map and two auxiliary maps for circular and elliptic mitochondria.

Main Results:

• The HED-Net effectively segments mitochondria of varied shapes and from complex backgrounds.
• The soft label-decomposition strategy enhances segmentation accuracy by exploiting shape knowledge.
• Extensive experiments on public benchmarks demonstrated that HED-Net outperforms existing state-of-the-art methods.

Conclusions:

• The proposed HED-Net provides a robust and accurate solution for semantic segmentation of mitochondria in EM images.
• The novel soft label-decomposition strategy improves the model's ability to handle diverse mitochondrial morphologies.
• This approach facilitates more reliable morphological analysis of mitochondria in biological research.