Comparison of EfficientNet CNN models for multi-label chest X-ray disease diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Chest X-ray analysis is crucial but challenging due to complexity and potential for human error.
• Existing deep learning models for chest X-ray diagnosis often have high computational costs and limited success rates.

Purpose of the Study:

• To develop a novel, lightweight, and fast deep learning architecture for multi-class autonomous detection of 14 diseases from chest X-ray images.
• To improve upon the performance and reduce the computational cost of previous chest X-ray diagnostic models.

Main Methods:

• Utilized the ChestX-Ray14 dataset (112,104 labeled images) for training and testing.
• Implemented and evaluated eight EfficientNet architectures (B0-B7) integrated with a coordinate attention mechanism.
• Proposed a hybrid, attention-based deep learning architecture.

Main Results:

• The EfficientNetB7 model achieved an AUC of 0.8265.
• The EfficientNet enhanced with coordinate attention achieved a higher AUC of 0.8309.
• The proposed model demonstrated superior performance compared to existing studies, particularly for eight of the 14 disease classes.

Conclusions:

• The developed lightweight and fast deep learning architecture effectively supports radiologists in disease diagnosis.
• Autonomous diagnostic systems utilizing this technology can enhance healthcare by preventing misdiagnoses and underdiagnoses.