Image-Text Neural Network for Patch-Wise CXR Pathology Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI algorithms for CXR abnormality analysis provide limited accuracy and reliability, particularly for pathologies like pneumothorax, and lack spatial information that aids medical practitioners in understanding and validating AI outputs.
Innovation Solution
An image-text deep neural network (DNN) processes medical imaging data using a vision transformer architecture, enabling patch-wise predictions of pathology presence or absence by correlating image and text inputs, with unsupervised learning and a two-stage training process to refine weights for enhanced accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If state-of-the-art AI algorithms provide classification results for overall CXR images, then the analysis can be performed automatically, but the output has limited use in clinical practice and it is difficult for medical practitioners to assess the relevance and reliability of the AI algorithm output
Solution Approach 1:
The patent divides the CXR image into multiple spatial patches and generates separate predictions for each patch. This segmentation allows the system to provide localized pathology detection results while maintaining automatic classification, enabling medical practitioners to assess the spatial distribution and relevance of findings in different regions of the image.
Solution Approach 2:
The patent transitions from providing only a global classification result to providing both global and local (patch-level) predictions. This adds a spatial dimension to the output, transforming the information structure from a single overall classification to a multi-level hierarchy that includes both aggregate and localized findings, thereby preserving and enhancing spatial information utility.
2Productivity
If reference AI algorithms are used for CXR abnormality analysis, then the analysis can be performed, but the accuracy is limited particularly for pathologies like pneumothorax that have varying manifestations across patients
Solution Approach 1:
The patent applies local quality by training the model to recognize pathology-specific visual characteristics in different spatial patches. Each patch prediction is tailored to detect local manifestations of pathologies, allowing the system to capture varying presentations of conditions like pneumothorax across different regions and patients, thereby improving detection accuracy while maintaining analysis capability.
Data Source
AI summary
Imaging data is processed in an image-text deep neural network, e.g., a vision transformer deep neural network. The image-text deep neural network also processes a text input indicative of a pathology. For each of multiple spatial patches within an observation region, a respective prediction of the presence or the absence of a finding of a pathology is provided.