Deep triplet hashing network for case-based medical image retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Medical Imaging

Background:

• Deep hashing methods excel at large-scale image retrieval but struggle with small-sample ranking in medical cases.
• Existing methods suffer from information loss related to classification, regions of interest (ROI), and limited samples, leading to inaccurate rankings.

Purpose of the Study:

• To propose an end-to-end framework, the Attention-based Triplet Hashing (ATH) network, for improved case-based medical image retrieval.
• To address the ranking problem in deep hashing by preserving crucial classification, ROI, and small-sample information.

Main Methods:

• Developed an Attention-based Triplet Hashing (ATH) network incorporating a spatial-attention module to focus on ROI information.
• Implemented a novel triplet cross-entropy loss function to maximize class-separability and hash code-discriminability, leveraging triplet labels for small-sample information.

Main Results:

• The ATH network effectively preserves classification, ROI, and small-sample information in learned hash codes.
• Extensive experiments on medical datasets show ATH outperforms state-of-the-art deep hashing methods in retrieval and ranking performance, particularly for small samples.
• The triplet cross-entropy loss significantly enhances classification performance and hash code-discriminability compared to other loss functions.

Conclusions:

• The proposed ATH network offers a robust solution for case-based medical image retrieval, overcoming limitations of existing deep hashing methods.
• ATH demonstrates superior performance in both general retrieval and small-sample ranking scenarios, highlighting the effectiveness of attention mechanisms and the triplet cross-entropy loss.