Medical Image Triage Using Two-Stage Normal Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based medical image reading systems face inefficiencies due to high workload for normal images and low sensitivity in distinguishing between weak normal and abnormal cases, leading to increased reader fatigue and reduced overall performance.
Innovation Solution
Implement a two-stage analysis using an abnormality prediction model to filter out 'strong normal' images with high sensitivity and a classification model to differentiate between 'weak normal' and abnormal images, reducing unnecessary reader workload by excluding 'strong normal' images from the worklist.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning based image recognition is used to automatically detect abnormalities in medical images, then detection capability is improved, but misclassification of normal images as abnormal increases
Solution Approach 1:
The patent introduces an intermediary filtering system between image acquisition and reader review. Normal images are automatically filtered out by AI-based abnormality detection, while only suspicious images are forwarded to radiologists. This intermediary layer prevents direct misclassification errors from reaching the final diagnostic workflow, thereby maintaining high detection capability while improving overall reliability through layered verification.
Solution Approach 2:
The system performs preliminary abnormality screening and classification before images reach the radiologist. By pre-processing images through automated anomaly detection and assigning priority levels, the system prepares images in advance for efficient human review. This preliminary action reduces the burden on radiologists and minimizes the impact of potential misclassifications by having multiple filtering stages ready beforehand.
2Reliability
If all medical images are reviewed by radiologists, then diagnostic accuracy is maintained, but work time and costs increase
Solution Approach 1:
The patent extracts and removes clearly normal images from the radiologist's review workload through automated AI-based classification. By identifying and filtering out normal cases with high confidence, the system extracts only the abnormal or suspicious images that require human expert review. This dramatically reduces radiologist work time while maintaining diagnostic accuracy, as radiologists focus exclusively on cases that need their expertise.
Solution Approach 2:
The system enables self-service processing where normal images are automatically handled by AI algorithms without requiring radiologist intervention. The automated triage system serves itself by routing images appropriately - normal images are processed and archived automatically, while only uncertain cases are escalated to human reviewers. This self-service approach minimizes time loss while preserving necessary human oversight.
3Productivity
If priority sorting of images is implemented, then work efficiency is improved, but additional processing steps are required
Solution Approach 1:
The patent implements priority sorting by changing the parameter of image classification from binary (normal/abnormal) to multi-level priority ratings. Images are assigned different priority levels based on the confidence and characteristics of AI detection results. This parameter change enables efficient workflow management where high-priority images are reviewed first, improving productivity without requiring complex additional hardware - only software-based classification layering.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A computing device comprising: a memory; and at least one processor that executes instructions stored in the memory, wherein the processor is configured to: obtain an abnormality score of a medical image; classify the medical image into normal or abnormal using the abnormality score; and add the medical image to a worklist required to be checked by a reader depending on a classification result.