Inverse Distance Map Boundary Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for delineating the boundary of an object in images, especially in medical imaging, require dense input and high-resolution images, making them costly and inefficient for accurate boundary computation, particularly when dealing with sparse image volumes or incomplete boundaries.
Innovation Solution
Training a convolutional neural network to predict an inverse distance map for boundary delineation, allowing for accurate boundary computation without relying on binary segmentation masks, and using a non-binary ground truth to focus on boundary representation rather than enclosed volumes, enabling the identification of partial or open boundaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep learning segmentation models are used to delineate object boundaries, then boundary delineation can be achieved, but dense input and high-resolution images are required, increasing computational cost
Solution Approach 1:
The patent extracts only the boundary information from the complete segmentation task. Instead of training the network to predict full segmentation masks, it trains the network to predict only boundary locations directly from sparse or low-resolution inputs, eliminating the need for dense input processing while maintaining boundary delineation accuracy
Solution Approach 2:
The patent inverts the conventional approach by not computing the enclosed volume first and then deriving the boundary. Instead, it directly predicts the boundary from the input images, reversing the traditional workflow of volume computation followed by boundary extraction
2Measurement precision
If conventional segmentation methods compute enclosed volume first, then boundary can be derived, but this approach requires complete enclosed space information which is unavailable for open boundaries
Solution Approach 1:
The patent extracts only boundary location information as the prediction target, discarding the requirement for complete enclosed volume computation. This extraction approach allows the model to handle both closed and open boundaries uniformly, as it only needs to identify boundary pixels without requiring knowledge of the enclosed space
Solution Approach 2:
The patent changes the prediction parameter from binary segmentation masks to continuous boundary location maps. By regressing boundary coordinates directly rather than classifying each pixel as inside/outside, the model can handle open boundaries and partial structures that cannot be represented as closed volumes
Data Source
AI summary
The disclosure relates to a method for determining a boundary about an area of interest in an image set. The includes obtaining the image set from an imaging modality and processing the image set in a convolutional neural network. The convolutional neural network is trained to perform the acts of predicting an inverse distance map for the actual boundary in the image set; and deriving the boundary from the inverse distance map. The disclosure also relates to a method of training a convolutional neural network for use in such a method, and a medical imaging arrangement.