Breast Microcalcification Diagnosis Using Deep Convolutional Neural Network from Digital Mammograms

Hongmin Cai, Qinjian Huang, Wentao Rong, Yan Song, Jiao Li, Jinhua Wang, Jiazhou Chen

TL;DR

This paper introduces a deep learning framework for automatically diagnosing breast microcalcifications (MCs) from mammograms, a key indicator of breast cancer. By combining deep learning features with traditional handcrafted features and employing a novel filtering strategy, the authors achieve improved classification accuracy, demonstrating the synergistic potential of these approaches.

What Problem Does This Paper Address?

Breast cancer is a leading cause of mortality among women, and early detection significantly improves survival rates. Mammography is a primary screening tool, and the presence of microcalcifications (MCs) is a crucial early sign of malignancy. However, accurately characterizing these MCs, which appear as tiny spots (0.1-1.0 mm), has been challenging. Traditional methods rely heavily on handcrafted image descriptors, which can be sensitive to variations in image quality and breast tissue composition, and lack a direct causal link to the diagnostic outcome. This paper aims to address the challenge of robust and automatic MC diagnosis by leveraging the power of deep learning, while also exploring how to best integrate it with established handcrafted features.

Methods

The authors propose a three-step methodology:

  1. ROIC Extraction: An automatic image preprocessing pipeline is used to extract the Region of Interest (ROI) containing suspicious calcifications. This involves morphological erosion, top-hat filtering, Otsu thresholding, and dilation to isolate potential MC areas.
  2. Feature Extraction:
    • Deep Features: A Convolutional Neural Network (CNN) based on the AlexNet architecture is employed. Pre-trained on ImageNet, the model is fine-tuned for mammography images to extract deep features from the penultimate layer.
  3. Handcrafted Features: A comprehensive set of traditional radiomics features are extracted, including first-order statistics, morphological features (200 descriptors), and texture features (352 descriptors).
  4. Classification and Fusion:
    • Individual Feature Sets: SVM classifiers are trained using deep features alone, handcrafted features alone, and filtered versions of these.
  5. Feature Fusion: Canonical Correlation Analysis (CCA) is used to fuse deep and handcrafted features. CCA aims to find linear combinations of features that maximize their correlation, thereby capturing complementary information.
  6. Filtered Deep Features: A novel approach involves filtering the deep features based on their correlation with handcrafted features, aiming to retain the most informative deep features.

Datasets

The study utilized a dataset comprising 990 full-field digital mammography images from two medical institutions: Sun Yat-Sen University Cancer Center (SYSUCC) and Nanhai Affiliated Hospital of Southern Medical University. These images were from 328 breast lesion cases, with 540 images showing malignant masses and 450 showing benign lesions, all confirmed by histopathology. Data augmentation through simple rotation was applied. The dataset was split into 3564 ROIs for training and 99 ROIs for testing.

Key Insights and Ideas

  • Deep Learning Outperforms Handcrafted Features: The study confirms that deep learning features, extracted via CNNs, generally provide superior classification performance compared to traditional handcrafted features. This highlights the ability of deep networks to learn complex and discriminative representations directly from image data.
  • Handcrafted Features Offer Complementary Information: Despite the success of deep learning, handcrafted features are not rendered obsolete. The paper demonstrates that combining deep and handcrafted features, or using handcrafted features to guide the selection of deep features, can further enhance diagnostic performance. This suggests that traditional features capture aspects of MCs that deep networks might miss or underemphasize.
  • Feature Filtering is Crucial: A key finding is that filtering deep features based on their correlation with handcrafted features leads to the best performance. This "filtered deep features" approach effectively prunes less relevant deep features, focusing on those that align with or complement established diagnostic indicators.
  • Visualization Reveals Diagnostic Clues: The visualization of neuron activations in the CNN shows that while some neurons capture traditional radiomic patterns, others highlight features like spicules and lobulation, which are associated with malignancy but not directly with calcifications themselves. This offers insights into how the CNN "sees" and interprets the image.

Key Results

The paper reports the following key quantitative results:

  • CNN Features: Achieved an accuracy of 0.8768 ± 0.0431.
  • Handcrafted Features (Morphological): Achieved an accuracy of 0.8525 ± 0.0203.
  • Combined Features (CNN + Morphological): Achieved an accuracy of 0.8828 ± 0.0437.
  • Filtered Deep Features (CNN filtered by Morphological): Achieved the highest accuracy of 0.8859 ± 0.0363, with a precision of 89.32% and sensitivity of 86.89%. This combination also yielded the best AUC (0.9392 ± 0.0240).

Limitations and Future Work

  • ROI Extraction: The authors acknowledge that the current ROI extraction is performed using an automatic processing method with some weaknesses. They propose that future work should focus on making both ROI extraction and lesion classification fully automatic and implemented by the CNN method.
  • Dataset Size: While the dataset of 990 images is substantial, the authors mention that the limited data necessitated data augmentation. Future work could benefit from larger, more diverse datasets.
  • Generalizability: The study was conducted on a specific dataset. Further experimental evaluation on different patient groups and diverse datasets is suggested to assess the generalizability of the proposed method.
  • Clinical Value: While the study demonstrates strong performance, further investigation into the clinical utility and interpretability of the deep features is warranted.

Summary updated

Abstract

Mammography is successfully used as an effective screening tool for cancer diagnosis. A calcification cluster on mammography is a primary sign of cancer. Early researches have proved the diagnostic value of the calcification, yet their performance is highly dependent on handcrafted image descriptors. Characterizing the calcification mammography in an automatic and robust way remains a challenge. In this paper, the calcification was characterized by descriptors obtained from deep learning and handcrafted descriptors. We compared the performances of different image feature sets on digital mammograms. The feature sets included the deep features alone, the handcrafted features, their combination, and the filtered deep features. Experimental results have demonstrated that the deep features outperform handcrafted features, but the handcrafted features can provide complementary information for deep features. We achieved a classification precision of 89.32% and sensitivity of 86.89% using the filtered deep features, which is the best performance among all the feature sets.
BibTeX
@inproceedings{cai2019breast,
  author = {Hongmin Cai and Qinjian Huang and Wentao Rong and Yan Song and Jiao Li and Jinhua Wang and Jiazhou Chen and Li Li},
  title = {{Breast Microcalcification Diagnosis Using Deep Convolutional Neural Network from Digital Mammograms}},
  booktitle = {Comput. Math. Methods Medicine},
  year = {2019},
  url = {https://openreview.net/forum?id=IGZV0NIN_VH},
}
Figure