[Paper Review] Automated non-mass enhancing lesion detection and segmentation in breast DCE-MRI

I. Álvarez, Javier Ramı́rez+3 more|arXiv (Cornell University)|Mar 12, 2018

Blind Source Separation Techniques1 citations

TL;DR

This paper proposes a novel CAD system for automated detection and segmentation of non-mass enhancing (NME) lesions in breast DCE-MRI using independent component analysis (ICA) to extract dynamic tissue characteristics and a support vector machine (SVM) for voxel-wise classification. By projecting new images onto an ICA-derived source space and optimizing the SVM hyperplane, the method achieves a DSC of 0.7215, significantly reducing false positives.

ABSTRACT

Non-mass enhancing lesions (NME) constitute a diagnostic challenge in dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) of the breast. Computer Aided Diagnosis (CAD) systems provide physicians with advanced tools for analysis, assessment and evaluation that have a significant impact on the diagnostic performance. Here, we propose a new approach for the specific problem of NME detection and segmentation, by taking advantage of independent component analysis (ICA) to extract a data-driven dynamic characterization of tissue. A set of independent sources was obtained from a dataset of patients, and the dynamic behavior of the different tissues was described by multiple dynamic curves, together with a set of eigenimages describing the scores for each voxel. A new test image is projected onto the independent source space using the unmixing matrix, and each voxel is classified by a support vector machine (SVM) that has already been trained with manually delineated data. A solution to the high false positive rate problem is proposed by controlling the SVM hyperplane location. The CAD system is trained and validated, reaching a DSC coefficient of 0.7215 for NME segmentation.

Motivation & Objective

• To address the diagnostic challenge posed by non-mass enhancing (NME) lesions in breast DCE-MRI, which are difficult to detect and segment using conventional methods.
• To develop a data-driven approach for characterizing dynamic tissue enhancement patterns in DCE-MRI using independent component analysis (ICA).
• To improve segmentation accuracy of NME lesions by leveraging dynamic curves and eigenimages derived from ICA decomposition.
• To reduce false positive rates in NME detection through adaptive SVM hyperplane control based on training data.
• To validate the proposed CAD system on clinical data, achieving high segmentation performance measured by the Dice Similarity Coefficient (DSC).

Proposed method

• Independent component analysis (ICA) is applied to a dataset of DCE-MRI scans to extract a set of independent sources representing dynamic tissue enhancement patterns.
• Each independent source is associated with a dynamic curve and an eigenimage that encodes the spatial scores of the corresponding tissue behavior across voxels.
• For a new test image, the unmixing matrix derived from training data projects the image into the independent source space, enabling dynamic tissue characterization per voxel.
• A support vector machine (SVM) classifier, trained on manually delineated NME lesions, performs voxel-wise classification based on the projected dynamic features.
• A novel hyperplane adjustment strategy is introduced to control the decision boundary and reduce false positive detections.
• The system is validated using a clinical dataset, with performance measured using the Dice Similarity Coefficient (DSC).

Experimental results

Research questions

• RQ1 Can ICA effectively extract dynamic tissue enhancement patterns from DCE-MRI data to support NME lesion characterization?
• RQ2 Can a data-driven dynamic representation derived from ICA improve the accuracy of NME lesion segmentation in DCE-MRI?
• RQ3 How does SVM-based classification on ICA-derived features compare to conventional CAD approaches in terms of segmentation performance and false positive rates?
• RQ4 Can hyperplane optimization in the SVM classifier significantly reduce false positive detections in NME lesion segmentation?
• RQ5 What is the achievable segmentation accuracy of the proposed CAD system, as measured by the Dice Similarity Coefficient (DSC)?

Key findings

• The proposed CAD system achieved a Dice Similarity Coefficient (DSC) of 0.7215 for NME lesion segmentation, indicating strong agreement with manual delineations.
• ICA successfully extracted multiple independent dynamic curves and eigenimages that effectively described tissue enhancement behavior across the breast.
• The integration of ICA-derived dynamic features with SVM classification improved lesion detection performance compared to baseline methods.
• The hyperplane control mechanism significantly reduced false positive rates, enhancing diagnostic reliability.
• The system demonstrated robust performance on clinical DCE-MRI data, validating its potential for clinical CAD applications.
• The method provides a data-driven, dynamic characterization of tissue that supports accurate and reproducible NME lesion segmentation.