A Technical Review of Convolutional Neural Network-Based Mammographic Breast Cancer Diagnosis

Lian Zou, Shaode Yu, Tiebao Meng, Zhicheng Zhang, Xiaokun Liang, Yaoqin XieComput. Math. Methods Medicine 2019 · DOI

TL;DR

This paper reviews the application of Convolutional Neural Networks (CNNs) for mammographic breast cancer diagnosis (MBCD). It categorizes existing CNN-based MBCD models into three groups: dedicated architectures, transfer learning approaches, and CNNs used as feature extractors, providing a technical overview of their pros and cons. The review highlights the current state, challenges, and future directions in this rapidly evolving field.

What Problem Does This Paper Address?

Early detection of breast cancer through mammography is crucial for improving patient outcomes. However, interpreting mammograms is challenging due to subtle visual differences between malignant and benign lesions, leading to potential misinterpretations, overdiagnosis, and unnecessary biopsies. While computer-aided diagnosis (CAD) systems have been developed for decades, the advent of deep learning, particularly CNNs, has shown immense promise in enhancing the accuracy and efficiency of MBCD. This paper addresses the need for a comprehensive technical review of how CNNs are being applied in this specific domain, consolidating the diverse approaches and identifying key trends and challenges.

Methods

The paper categorizes CNN-based MBCD models into three main groups:

  • Dedicated CNN Models: These involve designing novel CNN architectures or modifying existing ones. The key idea here is to create specialized networks tailored for MBCD tasks, often involving iterative optimization with medical instances.
  • Transferred CNN Models: This approach leverages pre-trained CNN models (typically trained on large natural image datasets like ImageNet) and fine-tunes them with mammographic data. This is particularly useful when medical datasets are limited, as it allows for faster training and better performance by transferring learned features.
  • CNN Models as Feature Extractors: In this paradigm, CNNs are used to extract high-level features from mammographic images. These extracted features are then fed into traditional machine learning classifiers (like SVMs) for the final diagnosis. This approach combines the representational power of CNNs with the interpretability of classical ML models.

The review also discusses techniques like data augmentation, transfer learning strategies (e.g., using pre-trained models as feature extractors or fine-tuning them), and information fusion (combining features from different sources or models).

Datasets

The paper lists several mammography databases used in the reviewed studies:

  • Public Databases:
    • DDSM (Digital Database for Screening Mammography): Described as the largest available resource.
  • BCDR-F03 (Breast Cancer Digital Repository): Contains film mammography (FM) images.
  • INbreast: A publicly accessible database.
  • MIAS (Mammographic Image Analysis Society): Contains digital images.
  • IRMA: Contains image patches from various datasets.
  • DREAM challenge: A large dataset (82,000 images) for online competition.
  • In-house Collections: Many studies utilize proprietary datasets, often collected by institutions.

The paper notes that most databases consist of digital mammography (DM) images, with BCDR-F03 being one of the few that includes FM images.

Key Insights and Ideas

  • Transfer Learning is Crucial: Given the often limited size of medical imaging datasets, transfer learning from models pre-trained on large natural image datasets (like ImageNet) is a dominant and effective strategy for improving MBCD performance.
  • CNNs Integrate Feature Extraction and Classification: Unlike traditional CAD systems that separate feature extraction and classification, CNNs can learn these processes end-to-end, leading to more powerful representations.
  • Hybrid Approaches Show Promise: Combining CNN-extracted features with traditional machine learning classifiers, or fusing information from multiple CNN architectures or different imaging views, can further enhance diagnostic accuracy.
  • Data Augmentation is Essential: Techniques like image rotation and flipping are vital for increasing the robustness of CNN models to variations in lesion presentation.
  • Interpretability Remains a Challenge: While CNNs achieve high performance, understanding why they make certain predictions (interpreting learned features) is still an active area of research.

Key Results

The paper summarizes performance metrics like AUC, ACC, SEN, and SPE from various studies. While specific aggregate results are not presented as a single figure, Table 1 shows that AUC values often range from 0.8 to 0.98, with ACC values also in a similar high range for many of the reviewed CNN-based models. The review highlights that CNN-based models, especially those employing transfer learning, generally outperform traditional ML-based models. For instance, studies comparing CNNs with ML models often show superior performance for CNNs.

Limitations and Future Work

Limitations:

  • Limited Medical Data: The scarcity of large, high-quality, and diverse mammographic datasets remains a significant bottleneck.
  • Technical Challenges: Issues like overfitting, selecting informative features from high-dimensional CNN outputs, and integrating multi-perspective information are ongoing challenges.
  • Generalizability: Ensuring models perform well across different imaging devices, patient demographics, and ethnic/genetic variations is crucial for real-world deployment.
  • Interpretability: The "black box" nature of deep learning models makes it difficult to understand their decision-making process, which is a barrier to clinical trust and adoption.
  • Focus on Mammography: This review is limited to mammographic imaging and does not cover other modalities like ultrasound or MRI.

Future Work:

  • Data Collection and Sharing: The paper emphasizes the need for more extensive data collection and sharing initiatives to build larger, more representative datasets.
  • Interpretability Research: Developing better methods for visualizing and understanding CNN features is critical for clinical acceptance.
  • Seamless Integration: Creating systems that can simultaneously detect, localize, and classify lesions in a single pipeline is a desirable future goal.
  • Clinical Translation: Bridging the gap between research findings and clinical practice through FDA approval and widespread adoption requires robust validation and demonstration of generalization.

Summary updated

BibTeX
@article{zou2019technical,
  author = {Lian Zou and Shaode Yu and Tiebao Meng and Zhicheng Zhang and Xiaokun Liang and Yaoqin Xie},
  title = {{A Technical Review of Convolutional Neural Network-Based Mammographic Breast Cancer Diagnosis}},
  journal = {Comput. Math. Methods Medicine},
  year = {2019},
  doi = {10.1155/2019/6509357},
}
Figure