Using Radiomics as Prior Knowledge for Thorax Disease Classification and Localization in Chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Radiomics

Background:

• Chest X-rays are crucial but manual interpretation leads to radiologist burnout and diagnostic delays.
• Radiomics offers quantitative feature extraction from medical images, showing promise for diagnostic assistance.
• The increasing volume of chest X-ray data necessitates automated analysis tools.

Purpose of the Study:

• To develop an end-to-end framework, ChexRadiNet, for improved chest X-ray abnormality classification using radiomics.
• To enhance disease detection and localization accuracy in chest X-rays.
• To provide an automated system that assists radiologists in interpreting medical images.

Main Methods:

• Developed ChexRadiNet, an end-to-end framework integrating radiomics features.
• Employed a lightweight triplet-attention mechanism for classification and abnormality highlighting.
• Utilized class activation maps to extract radiomic features, guiding robust image feature learning.

Main Results:

• ChexRadiNet demonstrated superior performance on three public datasets (NIH ChestX-ray, CheXpert, MIMIC-CXR).
• Achieved state-of-the-art results in disease detection (0.843 AUC) and localization (0.679 T(IoU) = 0.1).
• The framework converges to more accurate image region identification with radiomic feature integration.

Conclusions:

• ChexRadiNet effectively utilizes radiomics features to improve chest X-ray analysis.
• The developed framework offers a promising solution for automated disease detection and localization.
• Publicly available code aims to facilitate further development in automated radiological interpretation.