Pectoral Muscle Equalization in Digital Mammograms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The presence of the pectoral muscle in mammograms interferes with automated analysis algorithms, causing challenges in segmentation, visualization, and registration, and existing methods struggle with imperfect segmentation leading to suppression artifacts.
Innovation Solution
A predictive approach is adopted to create pectoral-muscle equalized images using manual outlines, which are then used to estimate pixel-level bias, allowing for flexible 'soft segmentation' and accurate image processing, incorporating image processing and machine learning techniques to suppress muscle bias and improve segmentation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pectoral muscle segmentation is performed using traditional algorithms, then the pectoral muscle can be identified and suppressed, but segmentation artifacts are introduced due to imperfect segmentation
Solution Approach 1:
The patent introduces a predictive model as an intermediary that generates a predicted pectoral muscle image, which serves as a mediator between the original mammogram and the final suppressed image. This predicted image acts as a soft mask that reduces harsh boundaries and segmentation artifacts while maintaining accurate pectoral muscle suppression.
Solution Approach 2:
The patent transforms the segmentation problem from a binary classification (muscle vs. non-muscle) into a continuous parameter estimation problem by predicting pixel intensity values. This parameter change from discrete segmentation to continuous intensity prediction eliminates hard boundaries and reduces segmentation artifacts.
2Measurement precision
If manual outlines are used to create pectoral-muscle equalized images, then segmentation accuracy is improved, but processing time and complexity increase
Solution Approach 1:
The patent performs preliminary action by using manual outlines to create training data for a predictive model. The model is trained offline on a dataset with manual annotations, and once trained, it can rapidly process new images without requiring manual intervention for each case, thus achieving high accuracy with fast processing.
Solution Approach 2:
The patent creates a predicted copy of the pectoral muscle appearance based on training data, which can be rapidly generated for new images. This copying approach allows the system to leverage manual annotation accuracy without repeating the time-consuming manual outlining process for each new image.
3Shape
If the pectoral muscle is suppressed using hard segmentation masks, then the muscle region is clearly defined, but suppression artifacts are introduced at boundaries
Solution Approach 1:
The patent changes the parameter representation from binary mask values (0 or 1) to continuous predicted intensity values. This allows for smooth transitions at boundaries and eliminates the harsh edges that cause suppression artifacts while maintaining clear pectoral muscle definition.
Solution Approach 2:
The patent applies different processing qualities to different regions: the predicted intensity values provide smooth, artifact-free suppression in boundary regions, while maintaining accurate pectoral muscle definition in the core muscle region. This local quality adjustment eliminates artifacts at critical boundaries.
Data Source
AI summary
Image analysis techniques applicable to mammograms and other types of images may include image normalization, image segmentation, forming a prediction bias image, and creating an equalized image based on the prediction bias image. Creation of the equalized image may include subtraction of the prediction bias image from the original image. Forming the prediction bias image may involve the use of trained predictors.