ABSTRACT:
The efficient manifold ranking (EMR) algorithm has been widely utilized in content-based image retrieval (CBIR). In this algorithm, each image is represented by low-level features that describe color, texture, and shape. However, low-level features have limitations in capturing semantic meaning. To enhance EMR performance in CBIR, this research proposes a fusion method called CoEMR.
CoEMR combines multi-rankings on low-level features with CNN features extracted from a CNN model to enhance the discriminative power of a query image compared to dataset images. Furthermore, CoEMR generates a similarity score between two input images, constructing a similarity learning model. Experiments demonstrate the effectiveness of the proposed methods in improving EMR quality. Additionally, the potential integration of CBIR with Large Language Models in Medical Image Diagnosis Systems is discussed.
Key words: Content-based medical image retrieval, Medical image, Efficient manifold ranking, Deep Metric Learning, Contrastive loss, Triplet loss, EMR learning, LLMs, Semantic similarity.