Breast cancer is among the leading causes of death in women aged 40 and above, with over 230,000 new cases in the US annually. Early detection dramatically improves survival rates, and microcalcifications (MCs) — tiny calcium deposits as small as 0.1 mm appearing as bright spots on mammograms — are among the earliest and most reliable indicators of malignancy. Their low contrast and small size make automated detection extremely challenging.
This project, conducted at the School of Electrical Engineering and Computer Science, Ohio University, developed a novel computer-aided detection (CAD) pipeline for MCs in digital mammograms. The work resulted in both a peer-reviewed publication and a US Patent Application.
The detection pipeline consists of four stages:
The raw grayscale mammogram is first segmented using Fuzzy C-Means (FCM) clustering to isolate the breast tissue region from the background, reducing the search space for subsequent processing and improving robustness to scanner-to-scanner intensity variations.
A morphological top-hat transform is applied to the segmented image to enhance small, bright, high-frequency structures (MCs) against the slowly varying breast tissue background. The transform effectively suppresses low-frequency luminance variation while preserving MC-scale features. A subsequent watershed transform precisely isolates the candidate ROI from the surrounding tissue.
Within the extracted ROI, Gibbs random field (GRF) clique patterns are used to model the statistical dependencies between neighboring pixels. Multi-pixel cliques (2-, 3-, and 4-pixel) are analyzed: a clique is flagged as a potential MC if its average intensity exceeds a threshold times the average intensity of the surrounding 5×5 neighborhood. Single-pixel cliques handle very small MCs embedded in high-intensity regions, requiring the central pixel to exceed the local 3×3 average plus a threshold offset.
The final detected MC candidates are marked on the original mammogram image, providing a visual overlay for clinical review.
| Metric | Value |
|---|---|
| Overall Detection Rate (DR) | 94.4% |
| Sensitivity | 93.7% |
| Detection Accuracy (DA) | 88.2% |
| False Negative Rate | 5.6% |
The algorithm efficiently reduces the search space to regions with high MC probability before running the GRF detector, making it computationally tractable for clinical deployment.
The core detection method was filed as a US patent application by Ohio University:
The patent extends the detection method to a network-delivered clinical decision support system: a server receives a medical image, runs the detection pipeline, and returns the evaluation to a clinician's device — enabling cloud-assisted mammography screening at scale.