Deep Learning-Based Radiomics of B-Mode Ultrasonography and Shear-Wave Elastography: Improved Performance in Breast Mass Classification

Xiang Zhang, Liang Ming, Zehong Yang, Chushan Zheng, Jiayi Wu, Bing Ou, Haojiang Li, Xiaoyan Wu, Baoming Luo, Jun ShenFrontiers in Oncology 2020 · DOI

TL;DR

This paper introduces deep learning-based radiomics signatures derived from both B-mode ultrasound (B-US) and shear-wave elastography (SWE) to improve the classification of breast masses. These novel radiomics signatures significantly outperform traditional methods like BI-RADS assessment and quantitative SWE parameters, offering a promising avenue for more accurate and reliable breast mass diagnosis.

What Problem Does This Paper Address?

Breast cancer diagnosis heavily relies on medical imaging, with ultrasound (US) being a valuable tool due to its non-invasiveness and lack of ionizing radiation. B-mode US, while sensitive, often suffers from low specificity, leading to unnecessary biopsies of benign lesions. Shear-wave elastography (SWE) offers an improvement by quantifying tissue stiffness, which correlates with malignancy, but its diagnostic performance still has limitations. The challenge lies in further enhancing the diagnostic specificity of US-based methods to accurately differentiate between benign and malignant breast masses, thereby improving patient management and reducing healthcare costs. This paper aims to address this by exploring the potential of deep learning-based radiomics, a technique that extracts complex quantitative features from images, to augment the diagnostic capabilities of both B-mode US and SWE.

Methods

The core of this study lies in the development and evaluation of two deep learning-based radiomics signatures: B-US-RS (from B-mode US images) and SWE-RS (from SWE images). The process involved several key steps:

  • Image Acquisition: B-mode US and SWE images were acquired from patients with US-visible solid breast masses.
  • Segmentation: Regions of interest (ROIs) representing the breast masses were manually segmented from both B-mode US and SWE images. This step was crucial for isolating the relevant information for feature extraction.
  • Deep Learning-Based Radiomic Feature Extraction: A convolutional neural network (CNN) architecture was employed to automatically extract a large number of quantitative radiomic features from the segmented images. This approach differs from traditional radiomics by learning informative representations directly from the data, reducing the need for manual feature engineering.
  • Feature Selection: To manage the high dimensionality of extracted features, a rigorous selection process was implemented. This included selecting features with high intra- and inter-rater reproducibility (using Dice similarity coefficient and ICC) and then employing the Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify the most discriminative features for classification.
  • Radiomics Signature Construction: The selected features were used to construct radiomics signatures (B-US-RS and SWE-RS) by calculating a "rad-score" based on a weighted combination of these features.
  • Performance Evaluation: The diagnostic performance of the developed radiomics signatures was evaluated using receiver operating characteristic (ROC) analysis. These performances were then compared against traditional methods, including radiologist assessment using the Breast Imaging Reporting and Data System (BI-RADS) and quantitative SWE parameters (Emax, Emean, Eratio, ESD). Statistical comparisons were made using the McNemar test.

Datasets

The study utilized a retrospective dataset of 291 women with US-visible solid breast masses from two centers. These patients were divided into three cohorts:

  • Training Cohort: 198 masses from center 1, used for developing the deep learning-based radiomics signatures.
  • Independent Validation Cohort: 65 masses from center 1, used to validate the performance of the developed signatures.
  • External Validation Cohort: 28 masses from center 2, used to assess the generalizability of the findings.

The paper does not introduce a new benchmark dataset but rather uses existing patient data collected between June 2015 and July 2017.

Key Insights and Ideas

  • Deep Learning's Power in Radiomics: The study highlights the effectiveness of deep learning in automatically extracting complex, high-dimensional radiomic features that are not easily discernible by the human eye. This moves beyond traditional handcrafted features.
  • Synergy of B-mode US and SWE: While both B-mode US and SWE have their strengths, the study demonstrates that deep learning can effectively leverage the information from both modalities to create powerful classification models.
  • Superiority over Existing Methods: The most significant insight is that the developed deep learning-based radiomics signatures (B-US-RS and SWE-RS) consistently outperformed established methods like BI-RADS assessment and quantitative SWE parameters across all validation cohorts. This suggests a paradigm shift in how ultrasound data can be utilized for breast mass classification.
  • Robustness Across Cohorts: The high performance of the radiomics signatures was maintained not only in the independent validation cohort but also in the external validation cohort, indicating good generalizability.

Key Results

The study reported exceptional performance for the deep learning-based radiomics signatures:

  • High Diagnostic Accuracy: In the training cohort, both B-US-RS and SWE-RS achieved an Area Under the Curve (AUC) of 0.99 (95% CI = 0.99–1.00). This performance was maintained in the independent validation cohort (AUC = 1.00) and the external validation cohort (AUC = 1.00).
  • Improved Specificity: B-US-RS and SWE-RS demonstrated significantly higher specificity compared to quantitative SWE parameters and BI-RADS assessment. For instance, in the training cohort, their specificity was 100% compared to 93% for BI-RADS.
  • Outperforming Quantitative SWE: The radiomics signatures also showed superior sensitivity and specificity compared to individual quantitative SWE parameters like Emax.
  • Clinical Significance: The study concluded that the integration of deep learning-based radiomics could significantly improve the classification ability of B-mode US and SWE for breast masses, potentially leading to more accurate diagnoses and better patient management.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-Vendor Equipment: The study used ultrasound equipment from a single vendor, which might limit the generalizability of the findings to other vendors. Future work should involve validation on multi-vendor datasets.
  • Manual Segmentation: The segmentation of breast lesions was performed manually, which is time-consuming and can introduce variability. The authors suggest that future research should focus on developing automatic segmentation methods using deep learning to streamline the process.
  • 2D SWE: Only two-dimensional SWE images were used. While the authors cite literature suggesting 2D SWE performance is comparable or even superior to 3D SWE in some contexts, exploring 3D SWE could be a future direction.

Despite these limitations, the study provides a strong foundation for using deep learning-based radiomics in breast ultrasound. Future work should focus on multicenter, multi-vendor validation and the development of automated segmentation techniques to further enhance the clinical applicability of these promising approaches.

Summary updated

Abstract

The B-US-RS and SWE-RS outperformed the quantitative SWE parameters and BI-RADS assessment for classifying breast masses. The integration of the deep learning-based radiomics approach would help improve the classification ability of B-mode US and SWE for breast masses.
BibTeX
@inproceedings{zhang2020deep,
  author = {Xiang Zhang and Liang Ming and Zehong Yang and Chushan Zheng and Jiayi Wu and Bing Ou and Haojiang Li and Xiaoyan Wu and Baoming Luo and Jun Shen},
  title = {{Deep Learning-Based Radiomics of B-Mode Ultrasonography and Shear-Wave Elastography: Improved Performance in Breast Mass Classification}},
  booktitle = {Frontiers in Oncology},
  year = {2020},
  doi = {10.3389/fonc.2020.01621},
}
Figure