Deep Neural Network Segmentation via Dual Representation Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image processing methods for detecting the collimation area in radiography images often result in false positive and false negative errors due to high variability in grayscale values, leading to incorrect segmentation and increased manual intervention.
Innovation Solution
A deep neural network is trained on two different representations of the segmentation target, a confidence level region map and a confidence level contour map, to improve segmentation performance by validating outputs through shape similarity algorithms, reducing errors and enhancing detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image processing methods are used for collimation area detection, then the detection process is simple and fast, but false positive and false negative errors occur frequently due to high variability in grayscale values
Solution Approach 1:
The patent divides the segmentation task into two distinct representations: a confidence level region map (pixel-wise probability of region presence) and a confidence level contour map (pixel-wise probability of contour presence). This segmentation allows the neural network to learn different aspects of the segmentation target separately, improving reliability by capturing both regional and boundary information independently.
Solution Approach 2:
The patent extends the segmentation output from a single representation to multiple representations by adding a new dimension. Instead of only predicting a region map, the system also generates a contour map, effectively moving from one-dimensional output to two-dimensional output space. This dimensional expansion enables the model to capture complementary information and reduces errors by validating across multiple representation types.
2Measurement precision
If deep neural networks are trained on multiple representations, then segmentation precision improves, but training time and computational resources increase
Solution Approach 1:
The patent combines multiple training tasks into a single unified neural network architecture. The network simultaneously learns to predict both the confidence level region map and the confidence level contour map through a shared backbone and multiple output heads. This merging approach allows the model to learn complementary representations in parallel during training, improving segmentation precision while avoiding the need for separate training processes that would multiply time consumption.
Data Source
AI summary
This invention is related to a method to improve the performance of a deep neural network (10) for the identification of a segmentation target (111) in a medical image (12, 110), comprising the steps of performing n training steps on said deep neural network (10) for the identification of said region of interest on two different representations (13, 14) of the same segmentation target (111), said representations (13,14) being definitions of the same segmentation target (111).