Sparse Annotation is Sufficient for Bootstrapping Dense Segmentation.

Area of Science:

• Neuroscience
• Computer Vision
• Machine Learning

Background:

• Accurate 3D reconstructions from biological imaging, especially brain neuropil, are crucial for understanding neural circuits.
• Instance segmentation for deep learning models requires extensive, labor-intensive ground-truth annotation data.
• Current methods for generating training data are time-consuming and require expert annotators.

Purpose of the Study:

• To develop a novel deep learning method for rapid 3D segmentation generation.
• To reduce the human effort and time required for creating training data for biological imaging.
• To enable non-expert annotators to contribute to large-scale neuroimaging data annotation.

Main Methods:

• Developed a deep learning-based approach to generate dense 3D segmentations from sparse 2D annotations.
• Utilized serial section electron microscopy data of brain neuropil.
• Trained models on rapidly generated segmentations and compared performance against expert annotations.

Main Results:

• Achieved rapid generation of dense 3D segmentations from minimal 2D annotations.
• Models trained on generated data showed comparable accuracy to those trained on expert ground-truth.
• Reduced human annotation time by three orders of magnitude, enabling non-expert contributions.

Conclusions:

• The novel deep learning method significantly accelerates the creation of training data for 3D instance segmentation.
• This approach democratizes the generation of large-scale training datasets for neuroscience research.
• Facilitates the study of brain circuits and the measurement of circuit strengths.