Mammogram CAD Using Multi-View CNN Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CAD systems for mammographic image analysis face challenges such as low repeatability in malignancy assessment due to subjective radiologist scores, false positives and negatives from 2D mapping of 3D breast structures, and reliance on single views or scales, which can lead to unnecessary procedures and stress for patients.
Innovation Solution
A computer-aided diagnostic method using a deep learning algorithm with a tailored Convolutional Neural Networks (CNN) model processes mammographic images from multiple views and scales to extract global and local features, generating a likelihood score for malignancy by combining these features and applying a Random Forest classifier for improved confidence and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a 3D breast structure is mapped onto a 2D plane by summing multiple images, then the mammographic image can be generated for X-ray examination, but false positives and false negatives occur due to overlapping contributions from multiple depth frames
Solution Approach 1:
The patent transitions from 2D mammographic images to 3D volumetric representation by stacking multiple 2D images from different depths to form a 3D breast model. This dimensional transformation allows the system to separate overlapping structures that appear conflated in 2D projections, thereby reducing false positives and negatives caused by image summation artifacts.
Solution Approach 2:
The patent divides the 3D breast volume into multiple discrete depth frames or slices. By segmenting the volumetric data into individual depth layers, the system can analyze each frame separately and identify abnormalities without the confounding overlap effects present in summed 2D images, thus improving diagnostic reliability.
2Device complexity
If existing CAD systems use information from only one view or one scale, then the system architecture is simplified, but false detections occur and diagnostic confidence is reduced
Solution Approach 1:
The patent combines information from multiple mammographic views (e.g., craniocaudal and mediolateral oblique views) and multiple scales (different magnifications or resolution levels) into a unified analysis framework. By merging these diverse data sources, the system achieves more robust and accurate detection while maintaining manageable system complexity through integrated processing architecture.
Solution Approach 2:
The patent creates a multi-functional CAD system capable of processing and analyzing mammographic data from various views and scales using a single unified algorithmic framework. This universal approach allows the same system to handle diverse input configurations without requiring separate specialized modules for each view or scale, thereby improving reliability without proportionally increasing complexity.
3Ease of manufacture
If a two-stage detection-classification framework is used to detect and classify abnormalities, then the processing approach is structured and systematic, but false positives occur when detected abnormalities are wrongly classified as malignant
Solution Approach 1:
The patent implements feedback mechanisms where the classification results are continuously refined based on additional analysis of the detected abnormalities. The system uses the output from the detection stage to guide further classification analysis, allowing for correction and refinement of initial classifications, thereby reducing false positives while maintaining the structured two-stage approach.
Solution Approach 2:
The patent performs preliminary analysis and filtering in the detection stage to prepare high-quality candidate abnormalities for classification. By pre-processing and pre-characterizing the detected regions with relevant features and context information before the classification stage, the system enables more accurate classification decisions and reduces the likelihood of wrong malignancy assignments.
Data Source
AI summary
A system and method for processing mammographic images of target breast tissue is provided. The mammographic images are processed to generate modified images. A deep learning algorithm, having a tailored Convolutional Neural Networks (CNN) model, is applied to the modified images to generate a first output and a second output. Global features associated with the entirety of the mammographic images are extracted by using the first output. Local features associated with Regions of Interest (ROIs) of the mammographic images are extracted by using the second output. The global features and the local features are combined and fuse to generate an indicator representative of likelihood of malignancy of the target breast tissue.