Medical Image Retrieval via Nearest Neighbor Search on Pre-trained Image Features.

Area of Science:

• Computer Science
• Medical Informatics
• Artificial Intelligence

Background:

• Nearest Neighbor Search (NNS) is crucial for high-dimensional data analysis, with applications in medical imaging, disease classification, and diagnosis.
• Traditional brute-force NNS is computationally infeasible for large medical datasets.
• Efficient NNS algorithms are vital for advancing medical image retrieval and analysis.

Purpose of the Study:

• To propose DenseLinkSearch (DLS), an efficient NNS algorithm for retrieving relevant images from heterogeneous medical image databases.
• To develop and evaluate a Transformer-based feature representation technique for medical image retrieval.
• To enhance the accuracy and speed of content-based medical image retrieval.

Main Methods:

• Developed DenseLinkSearch (DLS), an NNS algorithm that utilizes a pre-computed index of links for efficient database traversal.
• Proposed a novel Transformer-based feature representation method for medical images.
• Conducted extensive experiments comparing DLS and the proposed feature representation against state-of-the-art methods on benchmark and custom datasets.

Main Results:

• DLS demonstrated superior performance over existing NNS approaches in retrieval accuracy and speed.
• DLS achieved ≥99% R@10 on 11 out of 13 benchmark datasets, with lower average query times compared to approximate NNS methods.
• The proposed Transformer-based feature representation significantly outperformed existing pre-trained models, showing improvements of up to 13.33% in P@20.

Conclusions:

• DenseLinkSearch (DLS) offers an effective and efficient solution for nearest neighbor search in large-scale medical image databases.
• The novel Transformer-based feature representation enhances the capabilities of content-based medical image retrieval.
• The developed methods provide significant advancements for medical image analysis, diagnosis, and information retrieval.