Radon-Gabor barcodes for medical image retrieval

Mina Nouredanesh, Hamid R. Tizhoosh, Ershad Banijamali, James Tung

Introduction

Researchers have developed a novel method to transform complex medical X-rays into compact, searchable binary "barcodes" by fusing two powerful mathematical tools: the Radon and Gabor transforms. This approach achieves approximately 81% retrieval accuracy while remaining significantly more memory-efficient than traditional image-search methods.

The Challenge of Medical Image Retrieval

Modern healthcare facilities generate terabytes of medical imagery annually. Navigating these massive databases to find specific images for diagnosis or research is a significant challenge. Manual text tagging is slow, expensive, and often subjective.

Content-Based Image Retrieval (CBIR) offers a solution by searching based on visual features like shape and texture. However, traditional features often require substantial memory. For example, a "Bag of Words" approach for a million images might require 1GB of RAM, whereas binary embeddings (barcodes) can reduce that footprint to just 48MB. The challenge lies in creating a barcode that is compact enough for speed yet descriptive enough to distinguish between hundreds of different medical image classes.

Proposed Methodology: Fusing Radon and Gabor Transforms

The authors propose two novel ways to generate these barcodes by combining the Radon Transform, which excels at capturing global shapes and linear structures, and Gabor Filters, which are highly effective at capturing localized textures at various scales and orientations.

1. Gabor-of-Radon-Image Barcodes (GRIBCs)

In this approach, the Radon transform is first applied to the image to create a "shadowgram." A bank of Gabor filters is then applied to this shadowgram, and the results are binarized based on the median value to create the final barcode.

The GRIBC process: applying the Radon Transform (RT) to create a shadowgram, followed by Gabor filtering.

2. Guided-Radon-of-Gabor Barcodes (GRGBCs)

This "guided" method first applies Gabor filters to the original image to highlight specific textures. Subsequently, Radon projections are performed perpendicular to the orientation of the Gabor filter's stripes. This ensures the Radon transform focuses on the most meaningful structural information highlighted by the Gabor filters.

In both cases, the resulting binary strings can be compared using Hamming distance, an exceptionally fast calculation for modern computing systems.

Dataset and Experimental Setup

The study utilized the IRMA (Image Retrieval in Medical Applications) X-ray dataset, which includes 12,677 training images and 1,733 test images across 193 distinct classes. Images are annotated with a 13-character hierarchical code representing imaging modality, body orientation, anatomical region, and biological system.

Key Insights and Results

  • Shape and Texture Synergy: The authors demonstrate that while Radon captures the "skeleton" of an image, Gabor captures the "flesh" (texture). Combining them makes the retrieval system robust against noise and variations in scale or rotation.
  • Binarization Efficiency: Using the median value for thresholding ensures that exactly half the bits are '1' and half are '0', maximizing the information entropy of the barcode.
  • Accuracy: The best-performing method (GRGBC) achieved approximately 81% accuracy for the top search result.
  • Efficiency: The proposed Radon-Gabor combinations significantly outperformed simpler binary features like Local Binary Patterns (LBP) and standard Radon Barcodes (RBC).

Limitations and Future Directions

While binary codes are fast, the authors acknowledge that extremely long barcodes (some configurations reached over 8,000 bits) may eventually prolong retrieval time in massive archives. Future work should validate these barcodes on other types of medical imagery, such as MRI or CT scans, and address parameter sensitivity regarding the optimal number of scales and orientations for the Gabor filters.

Summary updated

BibTeX
@inproceedings{nouredanesh2016radon,
  author = {Mina Nouredanesh and Hamid R. Tizhoosh and Ershad Banijamali and James Tung},
  title = {{Radon-Gabor barcodes for medical image retrieval}},
  booktitle = {ICPR},
  year = {2016},
  doi = {10.1109/icpr.2016.7899818},
}
Figure