Abstract
DOAJ is a unique and extensive index of diverse open access journals from around the world, driven by a growing community, committed to ensuring quality content is freely available online for everyone.
Introduction
Mammography is the primary tool for early breast cancer detection, yet the visual environment of a mammogram is inherently "cluttered." Tumors are often subtle, low-contrast, and obscured by dense glandular tissue, making the screening process grueling for radiologists. Computer-Assisted Diagnosis (CAD) systems aim to mitigate this by automatically identifying Regions of Interest (ROI)—potential cancerous masses—before classification. This paper evaluates different image processing pipelines to optimize the accuracy and efficiency of this critical first step.
System Architecture
The identification of ROIs is part of a larger diagnostic pipeline. As shown in the flowchart below, the process begins with image acquisition and enhancement, followed by detection and feature extraction to facilitate classifier learning and performance analysis.
Methodology
The study compares two distinct stages of the image processing pipeline: enhancement and segmentation.
Image Enhancement: Fuzzy Logic vs. Histogram Equalization
The authors compared standard Histogram Equalization (HE) against Fuzzy Image Enhancement. While HE stretches global contrast, it often amplifies noise. The fuzzy approach maps pixel intensities into a "fuzzy plane" using membership functions, stretching contrast specifically around the mean intensity. This non-linear mapping preserves the contours of potential masses more effectively than HE, which can introduce noise artifacts.
Segmentation: Gradient-based vs. Region Growing
Two segmentation strategies were evaluated:
- Gradient-based techniques: Using operators like Sobel and Prewitt to identify edges where intensity changes abruptly. These were highly sensitive, capturing the actual mass in 96.7% of cases, but produced many false-positive regions.
- Region Growing (Clustering): This method starts with a "seed" pixel and expands to include neighboring pixels based on similarity. While using only the largest cluster caught the mass in 57% of cases, the authors suggest that using multiple clusters ($k > 1$) can match gradient sensitivity while significantly reducing the number of unimportant regions for subsequent analysis.
Experimental Results
The study utilized a subset of 30 mammograms from the University of South Florida (DDSM) database. Key findings include:
- Quantitative Superiority: Fuzzy enhancement consistently outperformed HE across metrics such as Distribution Separation Measure (DSM) and Target-to-Background Contrast (TBC).
- Clinical Benchmarks: Using the Sobel operator, the system achieved a Percentage Overlap (PO) of >50%—the clinical benchmark for success—in a significant portion of the test set. In several cases, the computer-generated ROI achieved a 100% overlap with expert radiologist sketches.
Limitations and Future Directions
The system currently struggles with "dense" breasts, where glandular tissue has similar X-ray attenuation to tumors. Future work will focus on embedding expert knowledge into the software to handle these subtle cases and expanding the feature set—incorporating texture and shape—to automatically prune false-positive ROIs that do not contain cancer.
Summary updated