Related Experiment Video
Updated: Aug 9, 2026
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
Human observer detection experiments with mammograms and power-law noise
A E Burgess1, F L Jacobson, P F Judy
1Radiology Department, Brigham and Women's Hospital, Harvard Medical School, Boston, Massachusetts 02115, USA. burgess@bwh.harvard.edu
Medical Physics
|May 8, 2001
Summary
This study determined contrast thresholds for detecting lesions of varying sizes in mammograms. Results show lesion size significantly impacts detection, with breast tissue acting as more than just random noise in certain tasks.
Area of Science:
- Medical Imaging
- Radiology
- Image Perception
Background:
- Mammography is crucial for early breast cancer detection.
- Understanding lesion visibility requires characterizing contrast thresholds.
- Mammographic backgrounds possess complex, non-stationary statistical properties.
Purpose of the Study:
- To determine contrast thresholds for lesion detection as a function of lesion size.
- To compare lesion detection performance in mammographic backgrounds versus filtered noise.
- To evaluate the applicability of statistical decision theory models to human performance in mammography.
Main Methods:
- Hybrid images were created by adding digital tumor images to digitized normal mammographic backgrounds.
- Lesion sizes ranged from 0.5 to 16 mm; contrast amplitudes were adjusted for 92% detection in two-alternative forced-choice (2AFC) and 90% in search tasks.
- Three observers (two physicists, one radiologist) participated in the experiments.
Main Results:
- In 2AFC tasks with mammographic backgrounds, threshold amplitudes increased with lesion size (>1 mm), indicating a contrast-detail diagram with a slope of 0.3.
- Human efficiency in 2AFC tasks on mammographic backgrounds was high (90% relative to models), suggesting non-random noise influences.
- In search tasks, performance was similar for mammographic and filtered noise backgrounds, implying breast structure acts as random noise for this task.
Conclusions:
- Mammographic backgrounds contain deterministic masking effects beyond random noise, impacting 2AFC lesion detection.
- Breast structure can be approximated as random noise for search-based lesion detection tasks.
- Statistical decision theory models remain valuable for estimating human performance in mammography, despite non-stationary background statistics.
