Curvilinear Object Segmentation via Deep Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing techniques face challenges in accurately and automatically segmenting curvilinear objects, such as blood vessels, from medical images, particularly in 2D and 3D scans, due to issues with boundary detection and noise, leading to false positives and fragmented boundaries.
Innovation Solution
A deep learning method using a trained deep neural network that learns latent representations of input images, determines boundary estimates, and generates segmentation maps, incorporating a recurrent neural network for context and correlation learning, and loss terms to encourage homogeneity and connectedness, while utilizing an edge prior for improved boundary localization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image processing techniques are used for segmentation, then the process is simple, but accuracy and reliability are poor due to noise and boundary detection issues
Solution Approach 1:
The patent replaces conventional mechanical image processing techniques with a deep learning-based neural network system. The neural network automatically learns latent representations and boundary estimates from input images, eliminating the need for manual feature engineering and conventional algorithms, thereby achieving higher reliability while managing complexity through automated learning
Solution Approach 2:
The patent introduces latent representations as an intermediary between input images and segmentation outputs. The neural network first learns compact latent representations that capture essential features, then uses these representations to generate boundary estimates and segmentation maps, effectively mediating the complex transformation from raw images to accurate segmentation
2Reliability
If deep learning methods are used to improve segmentation accuracy, then reliability improves, but the amount of training data required increases
Solution Approach 1:
The patent changes the approach to training by using unsupervised pre-training on large datasets to learn general latent representations, then fine-tuning on smaller annotated datasets. This parameter change in the training methodology allows the system to achieve high reliability while reducing the quantity of labeled training data needed, as the pre-trained model captures essential features without requiring extensive annotated examples
3Measurement precision
If boundary detection is performed to improve segmentation, then segmentation precision improves, but false positives and fragmented boundaries occur
Solution Approach 1:
The patent incorporates feedback mechanisms through loss functions that penalize false positives and fragmented boundaries. The neural network receives feedback during training about boundary accuracy, allowing it to iteratively improve its boundary detection. The loss function provides guidance to adjust boundary estimates, eliminating false positives and ensuring continuous, accurate boundaries in the final segmentation output
Data Source
AI summary
At least one input image comprising curvilinear features is received. Latent representations of the input images are learned using a trained deep neural network. At least one boundary estimate is determined based on the latent representations. At least one segmentation estimate of the at least one input image is determined based on the latent representations. The at least one image is mapped to output segmentation maps based on the segmentation estimate and the at least one boundary estimate.