Abstract
Efficient retrieval of similar medical images from large archives is a critical challenge in clinical decision support. This paper presents a two-stage approach combining Support Vector Machines (SVM) and Radon Barcodes (RBC). By classifying images into anatomical categories using Radon transform features and then performing a binary search within those categories, the system achieves high computational efficiency and a minimal memory footprint. Experimental results on the IRMA 2009 dataset demonstrate the feasibility of this geometric approach for large-scale medical image retrieval.
Introduction
Content-Based Image Retrieval (CBIR) in medicine aims to find "more cases like this" based on visual content rather than text-based metadata. A primary obstacle is the "semantic gap"—the difficulty in mapping low-level pixel data to high-level diagnostic meaning. Traditional methods often struggle with scalability or require specialized tuning for specific organs. This research proposes a generalized, efficient pipeline using the Radon transform to bridge this gap.
Methodology
The proposed system operates in two distinct phases: classification and retrieval.
Feature Extraction and Classification
The system utilizes the Radon transform, which captures the structural geometry of an image by taking projections at various angles. These projections are normalized and used as input for a multi-class SVM with a Radial Basis Function (RBF) kernel. The SVM's role is to narrow the search space by predicting the image's anatomical category (e.g., chest, cranium, pelvis) from 57 possible classes in the IRMA dataset.
Radon Barcodes (RBC)
To facilitate rapid searching, the Radon projections are binarized. By calculating a local median threshold for each projection angle, the image data is converted into a compact binary string, or "barcode." This binarization allows for extremely fast comparisons using Hamming distance.
Experimental Results
The method was evaluated using the IRMA 2009 dataset, consisting of 14,410 X-ray images.
- Classification Accuracy: The SVM achieved approximately 57% accuracy using 32 projection angles on $32 \times 32$ normalized images. While modest, this is competitive for the complex IRMA dataset using purely geometric features.
- Efficiency: The retrieval process is remarkably fast, averaging 51.5 milliseconds per query.
- Memory Usage: Each image is represented by a barcode of only 80 bytes. This efficiency allows for the indexing of massive datasets—up to one billion images—within just 80 GB of RAM.
Conclusion
The combination of SVM classification and Radon Barcodes provides a highly scalable solution for medical image retrieval. While classification accuracy remains a bottleneck, the efficiency and low memory requirements of the RBC approach make it a compelling candidate for real-time clinical applications. Future work may explore deep learning features to improve initial classification while retaining the speed of the barcode-based search.
Summary updated