Abstract: Digital mammography is a critical tool for breast cancer detection, yet images are often characterized by low contrast and significant noise. This paper introduces Fuzzy Contrast-Limited Adaptive Histogram Equalization (FCLAHE), a novel preprocessing filter that combines standard contrast enhancement with a non-linear fuzzy function. By incorporating spatial context through radial distance from a seed point, FCLAHE effectively suppresses background artifacts while preserving lesion details. Experimental results on 50 images from the DDSM database demonstrate a 14.16% improvement in segmentation accuracy compared to the industry-standard CLAHE method.
Introduction
Automated Computer-Aided Diagnosis (CAD) systems rely heavily on the quality of input images. Digital mammograms present a challenge due to their inherent intensity inhomogeneities and low signal-to-noise ratios. While Contrast-Limited Adaptive Histogram Equalization (CLAHE) is widely used to boost contrast, it frequently enhances background noise, leading to "ghost" lesions and increased false-positive rates in automated segmentation. The challenge lies in enhancing the lesion while simultaneously smoothing out distracting background noise.
The FCLAHE Method
The proposed FCLAHE filter addresses these limitations through a multi-step process. First, a standard CLAHE operation is applied for initial global contrast improvement. Subsequently, a "fuzzy" logic layer is introduced to handle local inhomogeneities. A seed point is identified at the brightest part of a suspected mass, and the algorithm calculates average intensities at various radial distances from that point.
A fuzzy membership function then compares individual pixels to their neighbors at the same radial distance. Pixels significantly brighter than the local average—indicative of noise rather than the natural gradual darkening of a mass—are assigned lower membership values and attenuated. This non-linear approach ensures that the background is smoothed without losing the vital internal characteristics of the tumor.
Experimental Results
The filter was evaluated using 50 mammographic images from the Digital Database for Screening Mammography (DDSM). Segmentation was performed using the Catarious Segmentation Method (CSM) and compared against expert-validated ground truths.
Quantitative analysis showed that FCLAHE significantly improved the overlap criterion (OV%). The average segmentation accuracy rose from 52% (without the filter) to 68.36% with FCLAHE, representing a 14.16% gain over standard CLAHE. Qualitatively, the filter successfully removed false microcalcifications in the background that were erroneously enhanced by standard methods.
Discussion and Conclusion
A key insight from the study is the transformation of the image histogram. FCLAHE converts the typically asymmetrical, spiky histogram of a raw mammogram into a smooth, symmetric Poisson-like distribution, which is more conducive to reliable segmentation. While the method currently requires manual seeding and can struggle with inhomogeneities physically attached to the mass boundary, it represents a significant step forward in preprocessing for CAD systems. Future work will focus on automating seed point selection and developing segmentation algorithms specifically optimized for the FCLAHE-transformed probability density function (PDF).
Summary updated