Texture analysis sorts out breast cancer types

2015 07 13 08 17 28 831 Scottish Flag Circle 200

Texture analysis of MRI breast images can reveal multiple lesion subtypes, potentially enabling different assessment and management strategies depending on lesion type, according to new Scottish-led research published in European Radiology.

In a study of 200 women with diagnosed primary breast cancer who subsequently underwent breast MRI, texture analysis of different lesion types showed that entropy-based features from the co-occurrence matrix (COM) had different internal enhancement patterns that were well correlated to lesion type. The differences may reflect different underlying growth patterns, and could be useful for treatment planning and assessment, the study team concluded.

In previous studies, lesion heterogeneity on imaging has been reported to be an important parameter in monitoring development and progression of disease, as well as response to therapy.

The study used texture analysis to identify whether underlying breast cancer subtypes can be classified based on pixel intensity distributions on MR images, wrote Dr. Shelley Waugh from Ninewells Hospital and Medical School in Dundee, U.K., along with colleagues from the University of Dundee and the MD Anderson Cancer Center in Houston, Texas (European Radiology, 12 June 2015).

2015 07 13 10 16 08 654 2015 07 10 Waugh 350
2015 07 13 10 16 07 267 2015 07 10 Waugh 800

Texture analysis and feature classification can be computationally intensive, and several steps are required in order to generate meaningful results from patient datasets. Images courtesy of Dr. Shelley Waugh.

"We have been investigating the utility of texture analysis in characterization of breast cancer on MRI, with a drive to understanding links between underlying pathology and appearances on imaging," Waugh wrote in an email to AuntMinnieEurope.com. "This work ... has demonstrated that the loss of estrogen receptor expression results in a more heterogeneous cancer, which could potentially have implications in monitoring neoadjuvant treatment response."

MRI use has grown as a complementary modality to mammography due to its high sensitivity and soft-tissue depiction, among other attributes. High volumes of MRI images have led to increased use of computer-based texture analysis (TA) -- a method of "statistically modeling spatial distributions of pixel gray levels in order to in order to recognize, classify, or segment the data," the study team wrote.

Classifying cancer subtypes

The study had two aims, first, to retrospectively create models using existing data to classify breast cancer subtypes, and second, to apply these models to create a blinded set of tests to evaluate model performance. The team first aimed to sort out the range of invasive cancers in terms of their immunohistochemical (IHC) profiles, which are seen clinically as hormone receptor (HR) positive (HR positive; HER2 negative), hybrid (HR positive, HER2 positive), HER2 overexpressed (HR negative, HER2 positive), and triple negative (TN; hormone and HER2 negative) lesion subtypes, the group wrote. They added that each category of lesion shows different growth patterns and requires different therapeutic approaches.

They looked at 200 women with newly diagnosed primary breast cancer who underwent pretreatment dynamic contrast-enhanced breast MRI performed on either a 16-channel 1.5-tesla scanner (Magnetom Avanto, Siemens Healthcare) using a two-channel breast matrix coil. Alternatively, imaging was performed on a 3-tesla Magnetom Trio scanner (Siemens Healthcare) with a 7-channel open breast biopsy coil.

Standard T1- and T2-weighted sequences followed by a 3D Fast Low Angle Short volumetric sequence were acquired axially. There were eight dynamic acquisitions in all, with two acquired before contrast injection and six following injection of 0.1 mmol/kg Dotarem (Guerbet) followed by a saline flush and maximum intensity projections (MIPs). The scanner software generated subtracted volumes and maximum intensity projection (MIP) images and DICOM images acquired after contrast.

Regarding the use of two field strengths in the study, a previous paper by the authors revealed no significant differences in the magnets used, and found that spatial resolution is the critical factor in textural analysis, Waugh explained.

Texture analysis: Training then study data

They performed texture analysis using MaZda version 4.7 software on square regions of interest of 100 pixels, placed across three sequential slices. The investigators created a model using retrospective training data, then applied it to test data, the authors wrote.

Next came gray-level normalization within the MaZda program to minimize the effects of brightness and contrast variations. COM features were used to perform texture analysis. Lesion subtypes were classified using a cross-validated k-nearest-neighbor technique, then accuracy relative to the pathology and calculated receiver operating curve (AUROC) was assessed. Finally, the investigators employed Mann-Whitney U and Kruskal-Wallis tests to assess the feature values of raw entropy.

This last step, feature values assessment, was aimed at determining whether raw feature values could be useful in differentiating between breast cancer subtypes, and also to quantify feature value differences between tumor subtypes and their relevance to tumor characteristics.

Histological subtyping accurate

The results showed that histological subtype classifications were similar across both training (148 cancers) and test sets (73 lesions) using the entire set of COM features, as follows:

  • Training: 75%, AUROC 0.816
  • Test: 72.5%, AUROC 0.823

The investigators found significant differences in entropy features between lobular and ductal cancers (p < 0.001). Still, entropy features alone were unable to create a robust classification model, they wrote.

"For all histological subtypes (ductal, lobular, and ductal carcinoma in situ) considered together, the classification accuracy was around 75%, however, the ROC value reflected excellent performance of the classification (0.816)," Waugh et al wrote. "Entropy features alone resulted in lower overall classification accuracies, but the raw entropy feature values were significantly different between lobular and ductal carcinomas.

Texture analysis performed well in the immunohistochemical profiling of lesions, with pair-wise classifications showing that HR+ cancers were distinct from HER2 and TN (i.e., HR-) cancers, the team wrote.

"This is in keeping with their different biological behaviors, such as aggressiveness," Waugh et al wrote. "The HR+ cancers also demonstrated significantly lower entropy values compared with the HR- cancers, which may be due to the tighter cell-cell junctions and differentiated epithelial monolayers."

New ground

The study is the first to prospectively identify whether breast cancer subtypes can be fully classified based on pixel density distributions on MR images, the authors wrote.

MR entropy-based features from the co-occurrence matrix, representing heterogeneity, provide critical information on tissue composition, the authors wrote. The study shows that entropy features can differentiate between both histological and immunohistochemical breast cancer subtypes. Different entropy features between breast cancer subtypes imply differences in lesion heterogeneity.

Textural differences on contrast-enhanced MR images may reflect underlying lesion subtypes, which provides additional information for clinical decision-making, the team concluded. Finally, differences between the lesion subtypes are worthy of being tested for different treatment responses, Waugh et al wrote.

With regard to estrogen receptors in the breast cancer tissue, assessed in this study via pathology, generally TNBC and HER2 cancers (i.e., those not expressing estrogen receptors) are more aggressive cancers, and "we have demonstrated a link in this study that the absence of the hormone receptor is associated with increased heterogeneity," in texture analysis, Waugh wrote. In this study, the estrogen receptors in the breast cancer tissue were assessed via pathology.

"While this is preliminary work, it provides the foundation for larger studies, potentially in line with a multicenter approach," she noted.

Page 1 of 19
Next Page