Unified Computational Intelligent Framework for Medical Image Smart Retrieval using Hybrid Feature Modeling

Abstract

Over the past few decades, Content-Based Image Retrieval (CBIR) has emerged as one of the most prominent research areas within the domain of computer vision. With the continuous growth of visual data and the rapid expansion of the internet, there is a growing demand for retrieval strategies that go beyond simple text-based methods and provide more relevant, user-oriented access to multimedia content. Various systems and tools have been developed to facilitate efficient querying and interaction with large-scale audio-visual databases. However, significant challenges remain, particularly when dealing with massive and diverse image repositories. In the medical field, digital imaging plays a crucial role, as images are generated from various instruments and modalities. The identification of image modalities is a critical step that helps narrow down the scope of analysis and enhances the precision of search results.

The objective of this research is to design a well-structured classification and retrieval framework capable of categorizing medical images based on their respective modalities. The proposed approach focuses on modality-based classification and retrieval of medical images. Experiments were conducted using the dataset employed for the ImageCLEF2012 modality classification task.

The classification process relies on  visual features  extracted from each image, including  Scale-Invariant Feature Transform (SIFT) ,  Local Binary Pattern (LBP) ,  Local Ternary Pattern (LTP) ,  Edge Histogram Descriptor (EHD) ,  Color and Edge Directivity Descriptor (CEDD) ,  Color Edge Detector using Wavelet Transform , and  Color Histogram . These features are subsequently combined into a single composite feature vector, which serves as input to a Support Vector Machine (SVM) classifier with a chi-square kernel for categorization into 31 distinct modality classes. The proposed framework achieved an overall classification accuracy of 72.2%, which represents a 2.6% improvement over the highest accuracy previously reported in the ImageCLEF2012 competition using visual feature-based methods.