Multi-scale Triplet Hashing for Medical Image Retrieval

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep hashing algorithms are crucial for efficient medical image retrieval in large datasets.
• Existing methods often overlook the importance of discriminate areas and hierarchical similarity in medical image features and hash codes.

Purpose of the Study:

• To develop a novel algorithm, Multi-scale Triplet Hashing (MTH), to address limitations in current medical image retrieval techniques.
• To effectively leverage multi-scale information, convolutional self-attention, and hierarchical similarity for improved hash code generation.

Main Methods:

• The MTH algorithm incorporates a multi-scale DenseBlock module to extract multi-scale image information.
• A convolutional self-attention mechanism is employed to capture discriminate areas by facilitating channel-wise information interaction.
• A novel loss function is designed to preserve category-level and semantic information while capturing hierarchical similarity.

Main Results:

• The MTH algorithm demonstrated superior performance in medical image retrieval tasks.
• Experiments were conducted on diverse datasets including X-ray, skin cancer, and COVID-19 radiography images.
• MTH significantly enhanced retrieval effectiveness compared to existing state-of-the-art algorithms.

Conclusions:

• The proposed MTH algorithm effectively addresses the limitations of previous deep hashing methods for medical image retrieval.
• Integrating multi-scale analysis, self-attention, and hierarchical similarity learning leads to more discriminative and effective hash codes.
• MTH offers a promising advancement for auxiliary diagnosis through improved medical image retrieval efficiency and accuracy.