Volume 41, Issue 8Part1 081904
Radiation imaging physics

Improving mass candidate detection in mammograms via feature maxima propagation and local feature selection

Jaime Melendez

Jaime Melendez

Diagnostic Image Analysis Group, Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, 6525 GA, The Netherlands

Author to whom correspondence should be addressed. Electronic mail: [email protected]; Telephone: +31 24 3652272; Fax: +31 24 3540866.

Search for more papers by this author
Clara I. Sánchez

Clara I. Sánchez

Diagnostic Image Analysis Group, Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, 6525 GA, The Netherlands

Search for more papers by this author
Bram van Ginneken

Bram van Ginneken

Diagnostic Image Analysis Group, Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, 6525 GA, The Netherlands

Search for more papers by this author
Nico Karssemeijer

Nico Karssemeijer

Diagnostic Image Analysis Group, Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, 6525 GA, The Netherlands

Search for more papers by this author
First published: 11 July 2014
Citations: 3

Abstract

Purpose:

Mass candidate detection is a crucial component of multistep computer-aided detection (CAD) systems. It is usually performed by combining several local features by means of a classifier. When these features are processed on a per-image-location basis (e.g., for each pixel), mismatching problems may arise while constructing feature vectors for classification, which is especially true when the behavior expected from the evaluated features is a peaked response due to the presence of a mass. In this study, two of these problems, consisting of maxima misalignment and differences of maxima spread, are identified and two solutions are proposed.

Methods:

The first proposed method, feature maxima propagation, reproduces feature maxima through their neighboring locations. The second method, local feature selection, combines different subsets of features for different feature vectors associated with image locations. Both methods are applied independently and together.

Results:

The proposed methods are included in a mammogram-based CAD system intended for mass detection in screening. Experiments are carried out with a database of 382 digital cases. Sensitivity is assessed at two sets of operating points. The first one is the interval of 3.5–15 false positives per image (FPs/image), which is typical for mass candidate detection. The second one is 1 FP/image, which allows to estimate the quality of the mass candidate detector's output for use in subsequent steps of the CAD system. The best results are obtained when the proposed methods are applied together. In that case, the mean sensitivity in the interval of 3.5–15 FPs/image significantly increases from 0.926 to 0.958 (p < 0.0002). At the lower rate of 1 FP/image, the mean sensitivity improves from 0.628 to 0.734 (p < 0.0002).

Conclusions:

Given the improved detection performance, the authors believe that the strategies proposed in this paper can render mass candidate detection approaches based on image location classification more robust to feature discrepancies and prove advantageous not only at the candidate detection level, but also at subsequent steps of a CAD system.