Related Experiment Video
Updated: Aug 24, 2025
07:30
Electromagnetic Navigation Transthoracic Nodule Localization for Minimally Invasive Thoracic Surgery
Published on: May 4, 2022
3.4K
Radiomics-Guided Global-Local Transformer for Weakly Supervised Pathology Localization in Chest X-Rays
IEEE Transactions on Medical Imaging
|October 26, 2022
Summary
Radiomics-Guided Transformer (RGT) accurately localizes and classifies cardiopulmonary pathologies in chest X-rays using only image-level labels. This novel approach integrates radiomic features without requiring bounding box annotations, improving diagnostic accuracy.
Area of Science:
- Medical Image Analysis
- Artificial Intelligence in Healthcare
- Radiomics and Deep Learning
Background:
- Traditional medical image analysis relied on handcrafted radiomic features, demanding precise pathology localization.
- Accurate localization is challenging in real-world settings, limiting the effectiveness of previous automated methods.
- Existing deep learning models for chest X-rays often overlook domain-specific radiomic features.
Purpose of the Study:
- To propose a novel Radiomics-Guided Transformer (RGT) for accurate cardiopulmonary pathology localization and classification.
- To develop a method that integrates global image context with local radiomic information without requiring bounding box annotations.
- To leverage image-level disease labels for end-to-end pathology localization and classification.
Main Methods:
- Developed a Radiomics-Guided Transformer (RGT) model comprising an image Transformer branch and a radiomics Transformer branch.
- Implemented fusion layers and cross-attention mechanisms to integrate image and radiomic features.
- Utilized a self-attention mechanism to extract bounding boxes for radiomic feature computation, creating a feedback loop for localization.
Main Results:
- RGT achieved superior performance in weakly supervised disease localization on the NIH ChestXRay dataset, outperforming prior methods by an average of 3.6% across IoU thresholds.
- Demonstrated improved classification accuracy, achieving 1.1% higher average area under the receiver operating characteristic curve.
- Successfully localized pathologies and classified diseases using only image-level labels, validating the end-to-end feedback loop.
Conclusions:
- The Radiomics-Guided Transformer (RGT) offers a powerful, annotation-efficient approach for medical image analysis.
- Integrating radiomic features guided by Transformer attention significantly enhances pathology localization and classification accuracy.
- The proposed method sets a new benchmark for weakly supervised learning in chest X-ray analysis.
