Abstract
Computer-Aided Diagnosis (CAD) systems are vital for screening mammography but often struggle with image noise and low-contrast boundaries. This paper presents a robust two-stage pre-processing pipeline designed to accurately isolate breast tissue by segmenting the breast contour and the pectoral muscle. By combining custom low-pass "integrator" masks with edge-preserving non-linear diffusion, the proposed method significantly improves the accuracy of segmentation—a critical first step for any CAD system. Evaluation on the MIAS database demonstrates that this approach outperforms traditional techniques in both boundary precision and noise reduction.
Introduction
The efficacy of CAD systems in radiology is frequently hampered by non-tissue artifacts and high-density regions like the pectoral muscle, which can mimic tumors and lead to false positives. Two primary challenges exist: defining the low-contrast breast boundary and excluding the pectoral muscle in Medio-Lateral Oblique (MLO) views. Existing methods often fail at the "ends" of the breast border where contrast is at its lowest or inadvertently blur critical boundaries. This study proposes a robust pre-processing framework to address these hurdles, ensuring a cleaner input for subsequent diagnostic algorithms.
Methodology
Breast Contour Extraction
The extraction process is designed to handle the low-frequency nature of the breast boundary. It begins with histogram equalization to normalize intensity across various mammograms, followed by a multi-step segmentation pipeline.
A $5 \times 5$ "integrator" mask is applied as a low-pass filter to smooth the image, facilitating effective thresholding. A novel contribution of this work is the specialized 1D mask designed to fix the top of the breast contour. Standard algorithms often fail here due to extremely low contrast. The authors identified that the real border usually sits in a "valley" of the intensity curve and designed a mask specifically to detect that local minimum.
Pectoral Muscle Segmentation
To segment the pectoral muscle without destroying its boundary, the authors employ non-linear (anisotropic) diffusion. Unlike standard Gaussian filters, this physics-based approach adjusts "diffusivity" based on the image gradient. This allows for noise reduction within the muscle region while actually sharpening the edge where the muscle meets the breast tissue. The process is restricted to a specific Region of Interest (ROI) defined by five control points derived from the previously extracted breast contour, ensuring computational efficiency.
Results and Discussion
The proposed methods were validated using 90 mammograms from the MIAS (Mammography Image Analysis Society) database, with expert radiologist annotations serving as the ground truth. Performance was measured using the Hausdorff Distance Measure (HDM) and the Mean of Absolute Error Distance Measure (MAEDM).
Quantitative results indicate that the "integrator mask" approach for contour extraction significantly outperformed polynomial modeling and active contour methods. Similarly, the non-linear diffusion technique for pectoral muscle segmentation yielded a mean error approximately half that of traditional Hough-Transform or Gabor-Filter methods. Qualitative analysis confirmed that the post-processing "valley fix" successfully mended broken contours in regions where other algorithms typically drop off.
Conclusion
By treating breast boundaries as low-frequency signals and utilizing edge-preserving diffusion for dense regions, this pre-processing pipeline provides a superior foundation for CAD systems. While the iterative nature of non-linear diffusion presents some computational overhead, the gain in segmentation accuracy is substantial. Future work will focus on testing the pipeline on larger, modern digital datasets to ensure generalizability across different imaging hardware.
Summary updated