Mitochondria Segmentation From EM Images via Hierarchical Structured Contextual Forest

Area of Science:

• Neuroscience
• Biomedical Imaging
• Computer Vision

Background:

• Mitochondrial morphology and distribution are vital indicators of neural dysfunction.
• Accurate delineation of mitochondria in electron microscopy (EM) images is challenging due to variations in appearance, size, shape, and complex surrounding structures.
• Leveraging contextual information is key to overcoming these segmentation challenges.

Purpose of the Study:

• To introduce a novel method for accurate mitochondria segmentation in EM images.
• To address the limitations of existing methods by incorporating extended contextual information.
• To develop an automatic and efficient approach for analyzing mitochondrial morphology and distribution.

Main Methods:

• Introduction of local patch pattern (LPP) features to reduce ambiguity in local appearance and texture.
• Proposal of an automatic segmentation method using iterative learning of a hierarchical structured contextual forest.
• Utilization of a novel median fusion strategy to enhance spatial and temporal contexts and suppress false detections.
• Extraction of LPP features from both images and historical predictions for a hierarchy of contextual features.
• Implementation of joint label prediction using structured random forest, avoiding computationally intensive graph-based methods.
• Development of a 2D segmentation variant with a hierarchical multi-view fusion strategy, maintaining accuracy.

Main Results:

• Achieved state-of-the-art performance with a Dice score of 90.9% on the public EPFL Hippocampus benchmark.
• Demonstrated the effectiveness of the proposed local patch pattern (LPP) features and iterative learning strategies.
• Quantitative comparisons confirmed the superiority of the developed method over existing approaches.
• The 2D variant achieved comparable accuracy to 3D segmentation, enhancing practical applicability.

Conclusions:

• The proposed method effectively segments mitochondria in EM images by utilizing novel contextual features and hierarchical structured prediction.
• The local patch pattern (LPP) features and median fusion strategy significantly improve segmentation accuracy and robustness.
• The developed approach offers a powerful tool for investigating mitochondrial dynamics in relation to neural health and disease.
• The method provides a state-of-the-art solution for automated mitochondria segmentation, advancing neuroscience research.