Topology Based Image Segmentation

Topological Data Analysis (TDA) extracts meaningful information from complex image data by extracting shape, structure, and other topological features. In this approach, the image is  transformed into a persistent homology topological representation that identifies key topological features that persist across different scales. An interactive interface allows an expert user to quickly differentiate between objects of interest and background clutter in these feature sets as part of an active learning framework. Once the data is sufficiently labeled, machine learning tools can be applied in parallel to provide high throughput processing of large datasets.

Topology Based Feature Extraction

Topological Data Analysis (TDA) extracts meaningful information from complex data by based on the shape, structure, and other topological features of data. This approach is particularly well suited for structured but noisy datasets where conventional image analysis approaches tend to perform poorly such as a typical plant root image.

RootShape segmentation in action.

Active Learning Approach to Automatic Segmentation

Ensuring segmentation accuracy is critical for many root segmentation applications. To ensure accurate and reliable performance,  GDA uses an active learning framework to allow an expert user to label a small subset of the feature space and then adaptively train the system to properly differentiate objects of interest from the background clutter. This interactive process allows the user to refine the training algorithm and high accuracy with confidence with minimal effort.

Active Learning

Cloud-based Scalable Processing

Once the active learning process is sufficiently trained to identify key components, multiple images can be processed parallel through a scalable, cloud-based processing framework. This eliminates the processing bottleneck and provides rapid, accurate, and consistent analysis of large sets of images.

GDA uses Topological Data Analytics techniques to extract critical features in minirhizotron and conventional images of roots. These features are then labeled through an active learning process informed by an expert user to differentiate between small subset of the roots and the surrounding materials. Once trained, the algorithm is able to rapidly, accurately, and consistently segments roots from large datasets.