A Topographic Representation for Mammogram Segmentation

Byung-Woo Hong, Michael Brady

Abstract

Breast cancer screening via mammography is a critical yet challenging task, with radiologists frequently missing subtle malignant masses. This paper introduces a novel segmentation framework that treats mammograms as 3D topographic surfaces. By analyzing the hierarchical nesting of intensity contours, the proposed method identifies salient anatomical regions and potential tumors in a contrast-invariant manner. The core of the approach is the "inclusion tree," which represents the global structural organization of the breast. Evaluation on the MIAS database shows a detection rate of 95.8%, highlighting the potential of topological analysis for robust Computer-Aided Diagnosis (CAD) systems.

Introduction

Mammography remains the primary tool for early breast cancer detection. However, the complexity of breast tissue and variations in imaging conditions make the identification of malignant masses notoriously difficult. While CAD systems have successfully addressed microcalcification detection, reliably segmenting masses remains an open challenge. Traditional methods often rely on local image features such as texture, edges, or statistical pixel distributions. These heuristics are frequently sensitive to noise and variations in image contrast across different patients and equipment.

To overcome these limitations, this research proposes a global, topology-based approach. Instead of focusing on local pixel variations, the method interprets the image as a topographic map where pixel intensity corresponds to elevation. This shift in perspective allows for a segmentation strategy that is inherently invariant to absolute intensity levels and focuses on the structural saliency of regions.

Methodology

The segmentation process involves four primary stages: noise reduction, topographic mapping, tree construction, and saliency measurement.

Pre-processing and Topographic Mapping

Raw mammograms are often characterized by noise that can lead to fragmented or jagged contours. To mitigate this, the image is first processed using an anisotropic diffusion filter. This technique smooths the image while preserving semantically significant edges.

Following filtration, the image is transformed into a topographic map using iso-level contours (isophotes), spaced uniformly between the image's minimum and maximum values. This representation effectively converts the 2D intensity field into a 3D surface model.

Topographic representation: (a) original intensity, (b) 3D surface interpretation, and (c) resulting iso-level contours.

The Inclusion Tree and Saliency

The central innovation of this work is the construction of an "inclusion tree." Based on the Jordan curve theorem, the algorithm determines the hierarchical relationship between contours—specifically, which contours enclose others. The image boundary serves as the root of the tree, with nested contours forming branches and leaves.

The inclusion tree concept: (a) nested regions in an image and (b) the corresponding hierarchical tree structure.

To identify suspicious regions, the algorithm calculates the Minimum Nesting Depth (MND). Salient features, such as tumors or the pectoral muscle, typically manifest as dense, quasi-concentric patterns of contours due to sharp intensity gradients. The MND counts the number of nested levels from a base contour to its innermost peak. High MND values serve as a robust indicator of structural saliency, allowing the system to distinguish significant masses from heterogeneous background tissue.

Experimental Results

The algorithm was evaluated using a subset of 48 mammograms from the Mammographic Image Analysis Society (MIAS) database. The test set included Mediolateral Oblique (MLO) views with masses of varying sizes and subtleties.

The topographic method successfully detected 46 out of 48 masses, achieving a 95.8% sensitivity. Qualitatively, the system demonstrated a unique ability to concurrently segment the breast boundary, the pectoral muscle, and candidate masses using a unified topological logic. This global parsing of the image structure ensures that the segmentation is contextually aware, rather than just a collection of isolated local detections.

Discussion and Conclusion

While highly effective, the system encountered two failure cases. One mass was missed due to extremely low contrast relative to dense surrounding tissue, which prevented the formation of a distinct nested structure. Another mass was missed because it was physically attached to the pectoral muscle without a clear intensity drop, causing the algorithm to merge the two structures.

Future work aims to leverage this topographic representation for broader clinical applications. The extracted dense regions can facilitate automated breast density measurement, a key biomarker for cancer risk. Furthermore, the segmented landmarks (boundary and muscle) provide a stable coordinate system for temporal image registration, allowing CAD systems to track changes in a patient's breast tissue over multiple years. By prioritizing topology over texture, this method provides a robust foundation for the next generation of mammographic analysis tools.

Summary updated

Abstract

This paper presents a novel segmentation method for delineating regions of interest (ROI’s) in mammograms. The algorithm concurrently detects the breast boundary, the pectoral muscle and dense regions that include candidate masses. The resulting segmentation constitutes an analysis of the global structure of the object in the mammogram. We propose a topographic representation called the iso-level contour map, in which a salient region forms a dense quasi-concentric pattern of contours. The topological and geometrical structure of the image is analysed using an inclusion tree that is a hierarchical representation of the enclosure relationships between contours. The “saliency” of the region is measured topologically as the minimum nesting depth. Experimental results demonstrate that the proposed method achieves a satisfactory performance as a prompt system in the mass detection.
BibTeX
@inproceedings{hong2003topographic,
  author = {Byung-Woo Hong and Michael Brady},
  title = {{A Topographic Representation for Mammogram Segmentation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2003},
  doi = {10.1007/978-3-540-39903-2\_89},
}
Figure