Related Experiment Video
Updated: Sep 11, 2025
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
Predicting the quality of digital mammography from the perspective of detectability of microcalcifications: A
Lucas E Soares1, Lucas R Borges2, Renato F Caron3
1Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, São Carlos, Brazil.
Background:
Image quality in mammography is influenced by several technical factors, including radiation dose, exposure parameters, and detector type. Although observer studies are essential for assessing image quality and lesion detectability, they are complex, time-consuming, and costly, making routine implementation challenging. Traditional objective metrics for quality assessment have limited applicability in medical imaging and often rely on a ground-truth reference, which is typically unavailable in clinical scenarios. Therefore, methods that assess image quality directly from clinical images, particularly when linked to diagnostic performance, are highly desirable.
Purpose:
To propose a method based on radiomic features and machine learning to predict the quality of clinical mammography images in terms of microcalcification detectability, without relying on routine observer studies.
Methods:
A total of 1125 radiomic features were extracted from clinical mammography images acquired using two different digital breast imaging systems. Feature selection was performed to reduce dimensionality and retain the most informative attributes. These were used to train a Multilayer Perceptron (MLP) regression model to estimate detectability values as defined by a model observer (MO). To increase the variability of the data and evaluate model robustness, two types of image degradation were synthetically applied to the dataset, one associated with noise and the other with blurring. The trained model was assessed using quantitative and statistical analyses, focusing on the correlation between predicted and reference detectability scores.
Results:
The proposed MLP regression model showed a strong correlation with the detectability values provided by the MO. Even with the introduction of degradations, the model maintained high predictive accuracy, achieving a correlation coefficient greater than 0.9.
Conclusions:
The combination of radiomic features and a machine learning regression model demonstrated the ability to account for variations in acquisition systems and simulated image degradations. This approach offers a promising tool for image quality assessment in clinical mammography, particularly in tasks involving microcalcification detection, without the need for labor-intensive observer studies.
