Sparse Annotation is Sufficient for Bootstrapping Dense Segmentation.

Area of Science:

• Neuroscience
• Computer Vision
• Bioimaging

Background:

• Accurate 3D reconstructions of neural circuits are crucial for understanding brain function.
• Deep learning models require extensive ground-truth data for training, which is labor-intensive to generate.
• Annotating instance segmentation in complex biological structures like the brain neuropil is particularly challenging.

Purpose of the Study:

• To develop a novel deep learning-based method for rapid generation of dense 3D segmentations from sparse 2D annotations.
• To reduce the human effort and time required for creating training data for biological image analysis.
• To enable non-expert annotators to contribute to the generation of training data for large-scale brain circuit mapping.

Main Methods:

• Developed a deep learning model to generate dense 3D segmentations from sparse 2D annotations on single serial section images.
• Utilized serial section electron microscopy data of the brain neuropil for method development and validation.
• Trained models using both rapidly generated segmentations and expert-annotated ground truth data.

Main Results:

• The novel method rapidly generates dense 3D segmentations from minimal 2D annotations.
• Deep learning models trained on these generated segmentations achieved accuracy comparable to models trained on expert-annotated data.
• Annotation time was reduced by three orders of magnitude, and non-experts could generate the required annotations.

Conclusions:

• The developed method significantly accelerates the creation of training data for 3D instance segmentation in biological imaging.
• This approach democratizes the generation of large-scale training datasets, facilitating brain circuit research.
• The findings pave the way for more efficient and accessible analysis of complex neural structures.