Masked Attention for 3D Image Patch Classification in CT Angiography
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer-aided detection systems for Pulmonary Embolism (PE) in CT angiography images face challenges in reducing false positive rates and effectively highlighting distinctive image regions, especially when these regions account for a small portion of the imaging data, due to interference from irrelevant image areas.
Innovation Solution
The proposed method combines soft attention with hard attention to generate attention masks that highlight distinctive image regions proportionally to their contribution while completely removing irrelevant regions, using a two-stage deep learning-based approach for candidate generation and classification, with a 3D U-Net model for segmentation and convolutional neural networks for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional CAD software is used for PE detection, then radiologists can detect and diagnose PE, but the false positive rate increases and interferes with radiologist reading
Solution Approach 1:
The patent segments the CT angiography volume into multiple 3D image patches, which are then processed individually through the attention mechanism. This segmentation allows the system to focus on specific regions of interest while filtering out irrelevant areas, thereby reducing false positives while maintaining detection accuracy.
Solution Approach 2:
The patent applies local quality by using attention masks that selectively enhance or suppress specific regions within each 3D image patch. The attention mechanism assigns different weights to different spatial locations, allowing the system to highlight distinctive image regions while suppressing irrelevant areas, thus reducing false positives without compromising overall detection reliability.
2Measurement precision
If attention mechanisms are applied to highlight distinctive regions, then detection accuracy improves, but irrelevant image regions may still cause confusion
Solution Approach 1:
The patent extracts and removes irrelevant image regions by applying attention masks that completely suppress certain areas. The masking mechanism identifies and extracts only the relevant distinctive regions while taking out (removing) the confusing irrelevant regions from the classification process, thereby improving measurement precision without introducing harmful confusion.
3Reliability
If small, irregular-shaped emboli are detected, then diagnostic accuracy improves, but detection becomes more difficult due to their small size and irregular shape
Solution Approach 1:
The patent addresses the detection difficulty of small, irregular-shaped emboli by transitioning to a 3D patch-based analysis approach. Instead of analyzing 2D slices individually, the system processes volumetric 3D patches that capture the spatial context and irregular shapes of emboli across multiple slices, making them easier to detect while maintaining high diagnostic accuracy.
Data Source
AI summary
A candidate generator generates a set of candidate three-dimensional image patches from an input volume. A candidate classifier classifies the set of candidate three-dimensional image patches as containing or not containing disease. Classifying the set of candidate three-dimensional image patches comprises generating an attention mask for each given candidate three-dimensional image patch within the set of candidate three-dimensional image patches to form a set of attention masks, applying the set of attention masks to the set of candidate three-dimensional image patches to form a set of masked image patches, and classifying the set of masked image patches as containing or not containing the disease. The candidate classifier applies soft attention and hard attention to the three-dimensional image patches such that distinctive image regions are highlighted proportionally to their contribution to classification while completely removing image regions that may cause confusion.