Chest X-ray Abnormality Detection Using ML Boundary and Vascular Indices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing chest X-ray image analysis techniques struggle to accurately detect abnormalities in areas where lungs and other organs overlap, such as the heart or liver, due to similar signal values and difficulty in setting precise boundary lines.
Innovation Solution
A method using machine learning to detect boundary lines of anatomical structures in chest X-ray images, setting third lung areas where lungs and other organs overlap, and extracting vascular indices to determine abnormal states based on reference indices from normal images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If density histogram and threshold-based methods are used to extract lung areas, then the method is simple and widely applicable, but the precision of boundary detection in overlapping areas (lungs with heart or liver) deteriorates
Solution Approach 1:
The patent changes the parameter used for boundary detection from simple density thresholds to machine learning-based classification. The system trains a classification model using features such as gradient direction, curvature, and texture patterns in the X-ray image, allowing accurate differentiation of lung boundaries even in overlapping areas where density alone is insufficient.
Solution Approach 2:
The patent introduces an intermediary classification model that acts as a mediator between the raw X-ray image and the final lung area extraction. This model processes intermediate features (edges, gradients, local patterns) to determine boundary locations, improving precision without requiring complex direct measurement methods.
2Productivity
If traditional image analysis methods are used to detect abnormalities, then the processing is fast and simple, but the detection accuracy in overlapping areas deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-training a classification model on labeled data containing normal and abnormal cases before actual abnormality detection. This pre-trained model can then quickly and accurately identify abnormalities in overlapping areas during clinical use, combining the speed of automated processing with the reliability of machine learning-based detection.
Solution Approach 2:
The patent replaces traditional mechanical image analysis methods (thresholding, edge detection algorithms) with a machine learning-based system. The neural network or classification model automatically learns optimal detection patterns from training data, substituting manual feature engineering and simple algorithms with an adaptive intelligent system that achieves higher accuracy.
Data Source
AI summary
A computer obtains a chest X-ray image, detects boundary lines in the chest X-ray image using a model constructed through machine learning, sets a third lung area including at least one of a first lung area or a second lung area, extracts a vascular index indicating at least one of thickness or density of at least one pulmonary blood vessel present in an area included in the third lung area, determines whether the area included in the third lung area is in an abnormal state on a basis of the vascular index and a reference index based on indices extracted in advance from an area in chest X-ray images in a normal state corresponding to the area included in the third lung area, and outputs, if determining that the area included in the third lung area is in an abnormal state, information indicating a result of the determination.