Spatial Matching of 2D Mammography Images and Specimen Radiographs: Towards Improved Characterization of Suspicious Microcalcifications

Noor Nakhaei, Chrysostomos Marasinou, Akinyinka Omigbodun, Nina Capiro, Bo Li, Anne Hoyt, William Hsu

Introduction

A significant challenge in breast cancer diagnostics is the "spatial gap" between radiological imaging and pathological findings. While mammography identifies suspicious lesions, the subsequent biopsy provides the definitive diagnosis. However, correlating the specific tissue analyzed by the pathologist back to its exact location on the original mammogram is difficult due to breast deformation during imaging and the small size of biopsy specimens. This research, presented by a team from UCLA, proposes an automated method to bridge this gap using microcalcifications as spatial landmarks.

The Clinical Challenge

When radiologists detect suspicious microcalcifications—tiny calcium deposits—the standard procedure is often a core needle biopsy. Despite their suspicious appearance, many of these calcifications are benign, leading to high false-positive rates, patient anxiety, and increased healthcare costs.

Clinical workflow for suspicious microcalcifications, highlighting the role of specimen radiographs in follow-up tests.

To improve AI-driven diagnostic tools, researchers must accurately correlate mammographic features with the "ground truth" of histopathology. The primary obstacle is that the breast is compressed during a mammogram, whereas the biopsy specimen is an uncompressed tissue cylinder. This makes pinpointing the tissue's origin on a 2D image a complex spatial registration problem.

Methodology: An Automated Pipeline

The researchers developed a multi-stage pipeline designed to use the stable spatial distribution of microcalcifications—likened to a "constellation" or fingerprint—to perform registration.

The proposed automated pipeline for spatially matching specimen radiographs to diagnostic mammograms using template matching.

The pipeline consists of four key stages:

  1. Segmentation: A two-stage algorithm isolates microcalcifications. Bright candidates are first identified via classical computer vision (Difference-of-Gaussians and Hessian analysis), then filtered by a convolutional regression model to ensure accuracy.
  2. Clustering: Using DBSCAN (Density-Based Spatial Clustering of Applications with Noise), the system groups related calcifications in the specimen radiograph. The largest cluster serves as the matching template.
  3. Template Matching: The mammogram is partitioned into a 300x300 non-overlapping grid. The specimen template is rotated (0°, 90°, 180°, and 270°) and compared against these patches using a cross-correlation metric.
  4. Scoring: Matches are identified based on the 99th percentile of similarity scores specific to each image, allowing the algorithm to adapt to varying cluster complexities.

Key Results and Insights

The study utilized a retrospective dataset of 80 patients, including magnified and standard mammographic views. Ground truth was established by expert radiologists.

The algorithm demonstrated high reliability in identifying non-matching regions, achieving a Negative Predictive Value (NPV) of 0.98 to 0.99 across various views. This high NPV acts as a clinical safety net; if the algorithm excludes a region, it is almost certainly not the source of the biopsy.

Metric Magnified ML Full CC Full ML/LM
Accuracy 0.99 0.99 0.99
Precision 0.66 0.67 0.69
Recall 0.61 0.58 0.63

While accuracy is high (largely due to the volume of true negatives in the grid), precision remains modest. Failures typically occurred when the biopsy needle altered the cluster's shape or when breast deformation was too extreme for 2D template matching.

Limitations and Future Directions

Current limitations include the use of a fixed grid, which can bisect clusters, and a limited set of tested rotations. Future work aims to implement more granular rotations, refine segmentation models for magnified views, and extend the system to handle cases with multiple biopsy sites in a single breast. By improving this spatial correlation, the researchers move closer to reducing unnecessary biopsies through more precise radiologic-pathologic correlation.

Summary updated

BibTeX
@inproceedings{nakhaei2021spatial,
  author = {Noor Nakhaei and Chrysostomos Marasinou and Akinyinka Omigbodun and Nina Capiro and Bo Li and Anne Hoyt and William Hsu},
  title = {{Spatial Matching of 2D Mammography Images and Specimen Radiographs: Towards Improved Characterization of Suspicious Microcalcifications}},
  booktitle = {Computer-Aided Diagnosis},
  year = {2021},
  doi = {10.1117/12.2582258},
}
Figure