Chest X-ray Lesion Localization via Deep CNN Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI intelligent auxiliary diagnosis methods for chest X-ray films are limited in interpretability, failing to accurately identify disease types corresponding to lesion areas and often produce high false positive rates due to poor-quality datasets and incorrect label extraction methods, leading to inconsistencies with clinical diagnoses.
Innovation Solution
An image recognition method using a deep convolutional neural network that preprocesses chest X-ray films, screens for posteroanterior chest images, and employs a binary classification model for negative and positive classification, followed by a detection model to accurately label disease types and outline lesion areas, utilizing a combination of Resnet-34 and fully-connected neural networks for feature extraction and localization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If slice-level disease labels are acquired from reports through natural language processing using internationally public datasets, then the AI intelligent auxiliary diagnosis system can be built, but the labels extracted from diagnostic reports cannot guarantee complete correctness, causing the trained model to deviate from actual doctor's diagnosis results
Solution Approach 1:
The system employs doctors to manually annotate lesion areas and disease types directly on X-ray images, replacing the unreliable automatic label extraction from diagnostic reports. This self-service approach by domain experts ensures high-quality training data that accurately reflects actual clinical diagnosis standards.
2Productivity
If whole-film disease classification is performed, then disease prediction can be made, but the method cannot specifically locate the disease position, making it inconsistent with clinical diagnosis requirements
Solution Approach 1:
The system segments the X-ray image into multiple candidate lesion areas using a detection model, then performs disease classification on each segmented region. This segmentation approach preserves spatial information about lesion positions while maintaining disease prediction capabilities, aligning with clinical diagnosis requirements that need both location and type information.
3Measurement precision
If class activation mapping method is used to achieve weak-supervised semantic segmentation, then approximate lesion area can be located, but the lesion area information is not accurate enough and has high false positive rate
Solution Approach 1:
The system introduces an intermediary detection model that generates multiple candidate lesion areas with confidence scores, then uses these candidates as input for subsequent classification. This intermediary step filters out low-confidence regions before classification, reducing false positives while maintaining accurate lesion localization.
4Measurement precision
If a detection model is used to detect disease type and label lesion area outline, then accurate lesion positioning is achieved, but the system complexity increases with multiple models required
Solution Approach 1:
The system merges the detection model and classification model into an integrated workflow where the detection model's candidate regions feed directly into the classification model. By combining these functions in a unified pipeline with shared feature extraction, the system achieves accurate lesion positioning and disease classification while managing overall system complexity through modular design.
Data Source
AI summary
A method includes the following steps: pre-processing chest X-ray films to obtain initial X-ray film images that meets format requirements; screening the initial X-ray film images to detect whether they are posteroanterior chest images; inputting the posteroanterior chest images into a binary classification model of the deep convolutional neural network for negative and positive classification; inputting the images presenting positive results into a detection model of the deep convolutional neural network to detect a disease type and label an outline of a lesion area in each image; and displaying the disease type and lesion area corresponding to the image.