Convolutional Network for Neural Structure Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reconstructing neural networks from electron microscopy images face challenges due to the small size and intricate branching of neurons, leading to errors in boundary detection and segmentation, and lack effective features for affinity graph generation, requiring significant human intervention and failing to achieve accurate image restoration.
Innovation Solution
Training a convolutional network with multiple layers of filters to produce images at the same or higher resolution than the original, using a combination of linear and non-linear filtering, and employing super-sampling to enhance image resolution and segmentation accuracy, while learning affinity graphs directly from unprocessed images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional segmentation methods are used, then the process is simple to implement, but the manufacturing precision deteriorates due to small size and intricate branching of neurons causing boundary detection errors
Solution Approach 1:
The patent divides the segmentation process into multiple stages: initial rough segmentation using conventional methods, followed by iterative refinement using trained filters that progressively improve boundary accuracy. Each filter stage segments the image at a different level of detail, allowing simple methods to handle easy cases while complex methods refine difficult boundaries.
Solution Approach 2:
The patent performs preliminary training of filters using labeled training images before actual segmentation. This preliminary action creates pre-trained filters that capture edge and boundary characteristics specific to neural tissue, enabling subsequent segmentation to achieve high precision without manual intervention during the actual segmentation process.
2Device complexity
If hand-designed filtering architectures are used, then the device complexity is low, but the measurement precision deteriorates due to lack of effective features for affinity graph generation
Solution Approach 1:
The patent transforms fixed hand-designed filter parameters into learnable parameters through training on labeled data. The filters adapt their weights and characteristics based on the specific features of neural tissue in the training set, allowing the system to automatically discover effective features for affinity graph generation rather than relying on generic hand-designed architectures.
Solution Approach 2:
The patent enables the filtering architecture to self-optimize by training on labeled images and automatically adjusting filter parameters. The system serves itself by learning from training data what features are most important for segmentation, eliminating the need for manual design of complex filtering architectures while achieving superior precision.
3Reliability
If human intervention is used, then the reliability of segmentation is high, but the productivity deteriorates due to weeks to months of human effort required for proofreading
Solution Approach 1:
The patent implements automated quality control through trained filters that independently verify and refine segmentation results without human intervention. The system performs self-correction by comparing against learned patterns from training data, maintaining high reliability while eliminating the need for slow manual proofreading processes.
Solution Approach 2:
The patent implements iterative refinement where trained filters continuously process and improve segmentation results in an automated pipeline. This continuous automated processing maintains high reliability through multiple refinement passes while achieving productivity gains by eliminating interruptions for manual review, processing images continuously without human intervention cycles.
4Measurement precision
If image resolution is increased through super-sampling, then the measurement precision improves, but the use of energy deteriorates due to increased computational requirements
Solution Approach 1:
The patent applies super-sampling selectively only to regions containing neural structures of interest rather than uniformly to the entire image. By segmenting the image into regions of interest and applying high-resolution processing only where needed, the system achieves high measurement precision for critical areas while reducing overall computational energy consumption compared to full-image super-sampling.
Data Source
AI summary
Identifying objects in images is a difficult problem, particularly in cases an original image is noisy or has areas narrow in color or grayscale gradient. A technique employing a convolutional network has been identified to identify objects in such images in an automated and rapid manner. One example embodiment trains a convolutional network including multiple layers of filters. The filters are trained by learning and are arranged in successive layers and produce images having at least a same resolution as an original image. The filters are trained as a function of the original image or a desired image labeling; the image labels of objects identified in the original image are reported and may be used for segmentation. The technique can be applied to images of neural circuitry or electron microscopy, for example. The same technique can also be applied to correction of photographs or videos.