Ultrasound Autoencoder Shadow Detection via Beta Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for image processing of ultrasound images face challenges in accurately determining shadows, particularly due to non-uniform training data and supervised learning methods, which can lead to inconsistent determination standards and suboptimal learning outcomes.
Innovation Solution
A learning model based on an autoencoder with a shade decoder and a subject decoder, utilizing a combining function to synthesize outputs and employing likelihood functions based on beta distributions to inhibit pixel values from concentrating at extremes, enabling unsupervised learning and accurate shadow detection in ultrasound images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with manual labeling is used to detect shadows in ultrasound images, then training data can be obtained, but determination standards become non-uniform and learning quality deteriorates
Solution Approach 1:
The system performs unsupervised learning where the model automatically learns shadow patterns from unlabeled ultrasound images without manual annotation. The autoencoder structure enables the system to self-organize and identify shadow regions through reconstruction error analysis, eliminating the need for human-labeled training data and ensuring uniform determination standards across all images.
Solution Approach 2:
Instead of directly detecting shadows through supervised classification, the patent inverts the approach by using an autoencoder to reconstruct images and identifying shadows through reconstruction errors. Regions with high reconstruction errors indicate shadow areas, transforming a classification problem into a reconstruction-based detection method that avoids manual labeling inconsistencies.
2Manufacturing precision
If beta distribution likelihood functions are used to constrain pixel values, then pixel concentration at extremes is prevented, but computational complexity increases
Solution Approach 1:
The patent changes the parameter distribution assumption from uniform or normal distributions to beta distributions for modeling pixel intensities. The beta distribution parameters are optimized during training to reflect the actual intensity distribution in ultrasound images, preventing pixel values from concentrating at minimum or maximum extremes and improving reconstruction quality.
Solution Approach 2:
The loss function incorporates beta distribution likelihood terms that provide feedback during training to constrain pixel value distributions. The gradient from the beta distribution likelihood guides parameter updates to maintain realistic intensity distributions, creating a feedback mechanism that prevents extreme value concentration without requiring complex external constraints.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A learning device (10) is configured to input an output from an encoder to which an input image is input to a shade decoder and a subject decoder. The learning device (10) is configured to execute learning of the encoder, the shade decoder and the subject decoder, based on a reconstruction error, an NLL for a shade image relating to shades in ultrasound images, and an NLL for a subject image relating to subjects in ultrasound images. The reconstruction error is an error between the input image and an output image obtained by using a combining function for synthesizing the shade image and the subject image.