Combining Deep Learning and Knowledge-driven Reasoning for Chest X-Ray Findings Detection.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Radiology Informatics

Background:

• Deep learning models are increasingly used in medical imaging but often ignore valuable information in radiology reports.
• Integrating knowledge from radiology reports can significantly boost the performance of deep learning algorithms.

Purpose of the Study:

• To develop a hybrid framework combining deep learning with radiology report knowledge for improved medical image analysis.
• To enhance chest X-ray (CXR) finding detection by leveraging unstructured text data from reports.

Main Methods:

• Utilized a comprehensive chest X-ray findings vocabulary for automatic annotation of X-rays using radiology reports.
• Developed a vocabulary-driven concept annotation algorithm for data preparation.
• Trained a deep neural network classifier on annotated X-rays.
• Implemented a knowledge-driven reasoning algorithm to refine deep learning performance.

Main Results:

• The annotated X-ray dataset facilitated the training of a deep neural network for finding detection.
• The knowledge-driven reasoning algorithm successfully improved the deep learning module's accuracy.
• A hybrid approach combining deep learning and radiology report knowledge demonstrated superior performance in CXR finding detection.

Conclusions:

• Integrating knowledge from radiology reports into deep learning models offers a significant advantage for medical image analysis.
• This hybrid framework shows promise for enhancing diagnostic accuracy in chest X-ray interpretation.