Automatic deep learning-based normalization of breast dynamic contrast-enhanced magnetic resonance images

Jun Zhang, Ashirbani Saha, Brian J. Soher, Maciej A. MazurowskiarXiv 2018 · arXiv

Abstract

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is a cornerstone of breast cancer diagnosis and radiogenomics. However, intensity inconsistencies across different scanners and acquisition parameters hinder the reliability of quantitative radiomic features. We propose a fully automatic, parameter-free normalization method that leverages deep learning-based tissue segmentation and piecewise linear mapping. By standardizing the intensities of key biological "anchors"—air, fat, dense tissue, and the heart—our method removes scanner-induced variations. Evaluation on a dataset of 460 subjects shows that normalized images yield more consistent radiomic features and improved predictive performance for tumor subtyping.

Introduction

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is vital for breast cancer prognosis and radiogenomics. Despite its utility, images acquired from different scanners or with varying parameters (e.g., echo time, repetition time, and magnet strength) exhibit significant intensity range discrepancies. These inconsistencies make it difficult to extract comparable radiomic features across diverse datasets. Simple normalization techniques like min-max scaling often fail due to anatomical differences, while complex physical modeling is error-prone and requires unavailable scanner metadata. This study introduces a robust, data-driven normalization pipeline to address these challenges.

Methods

The proposed normalization pipeline consists of two primary steps: tissue segmentation and intensity mapping.

Tissue Segmentation

The method utilizes two Fully Convolutional Networks (FCNs) based on the U-Net architecture to identify biological "anchors":

  1. Heart Segmentation: An FCN segments the heart from post-contrast images, where it is most visible.
  2. Breast and Tissue Segmentation: A second FCN segments the breast region from pre-contrast images, further distinguishing dense fibroglandular tissue from fat.
  3. Air Detection: Air is heuristically defined as the darkest 5% of voxels.

Piecewise Linear Mapping

After segmenting these four regions, the median intensity values for each tissue type are calculated. A subject-specific piecewise linear transformation then maps these values to those of a reference "archetype" patient profile derived from the training set. This process ensures that identical tissue types across different patients occupy a consistent intensity space.

Overview of the two-step normalization pipeline: tissue segmentation followed by piecewise linear mapping.

Dataset

The study utilized a clinical DCE-MRI dataset of 460 breast cancer subjects, scanned at 1.5T or 3.0T using GE and Siemens equipment. The data was split into 360 subjects for training the segmentation models and defining the archetype, with 100 subjects reserved for testing.

Results

The normalization method demonstrated significant improvements in both feature consistency and predictive utility.

Feature Standardization

Prior to normalization, 15 evaluated radiomic features showed high dependency on scanner parameters (TE, TR, and magnet strength). Post-normalization, 14 of these features became independent of these parameters, indicating successful standardization across different acquisition environments.

Visual and Predictive Improvements

Visual inspection confirms that the normalization process harmonizes image contrast and brightness across subjects. Furthermore, in a radiogenomic task predicting the Luminal A tumor molecular subtype, the Area Under the Curve (AUC) improved for 10 out of 15 features when using normalized images.

Comparison of breast DCE-MRI images before and after normalization for three different subjects, showing improved consistency in intensity and contrast.

Computational Efficiency

The entire process is highly efficient, requiring less than 30 seconds per subject on a single Nvidia GTX 1080 GPU, making it suitable for integration into clinical workflows.

Discussion and Conclusion

While the method effectively standardizes overall intensities, it does not currently account for varying noise characteristics between scanners. Future work could incorporate denoising steps to further refine features sensitive to image noise, such as the Signal Enhancement Ratio (SER). Additionally, expanding the evaluation to include non-axial views and a wider variety of scanner models would enhance the tool's generalizability.

In conclusion, this deep learning-based normalization approach provides a robust, parameter-free solution for standardizing breast DCE-MRI intensities, thereby unlocking the full potential of multi-center radiomic and radiogenomic studies.

Summary updated

BibTeX
@misc{zhang2018automatic,
  author = {Jun Zhang and Ashirbani Saha and Brian J. Soher and Maciej A. Mazurowski},
  title = {{Automatic deep learning-based normalization of breast dynamic contrast-enhanced magnetic resonance images}},
  year = {2018},
  eprint = {1807.02152},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.02152},
}
Figure