Precision medical image hash retrieval by interpretability and feature fusion

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Medical image retrieval accuracy is often limited by high inter-class similarity and lesion omission.
• Chest X-ray analysis presents challenges due to subtle visual differences and potential for missed findings.

Purpose of the Study:

• To develop a precise medical image hash retrieval method that combines interpretability and feature fusion.
• To enhance the accuracy of medical image retrieval, specifically for chest X-rays.

Main Methods:

• Utilized DenseNet-121 pre-trained with the comparison to learn (C2L) method for robust medical feature extraction.
• Integrated an interpretable saliency map for attention-based region localization and feature fusion.
• Employed a hash layer with classification, quantization, and bit-balanced loss functions for high-quality hash code generation.

Main Results:

• The interpretable saliency map effectively identified focal regions in chest X-rays.
• Feature fusion successfully prevented information omission, leading to more comprehensive data representation.
• The combined loss functions generated accurate hash codes, improving retrieval accuracy over existing methods.

Conclusions:

• The proposed hash retrieval method enhances medical image retrieval accuracy through interpretability and feature fusion.
• This approach shows potential for integration into computer-aided diagnosis systems for improved clinical decision-making.