Preoperative Non-Invasive Prediction of Breast Cancer Molecular Subtypes With a Deep Convolutional Neural Network on Ultrasound Images

Chunxiao Li, Haibo Huang, Ying Chen, Sihui Shao, Jing Chen, Rong Wu, Qi ZhangFrontiers in Oncology 2022 · DOI

TL;DR

Researchers have developed a deep convolutional neural network (DCNN) that predicts breast cancer molecular subtypes directly from preoperative ultrasound images and basic clinical data. By outperforming traditional radiologist-led lexicon models, this approach offers a non-invasive path toward personalized treatment planning before a patient ever enters the operating room.

What Problem Does This Paper Address?

Breast cancer is not a monolithic disease; it is categorized into four molecular subtypes (Luminal A, Luminal B, HER2-enriched, and Triple-Negative) based on hormone receptors and genetic markers. Each subtype carries a vastly different prognosis and requires a specific treatment strategy—ranging from endocrine therapy to aggressive chemotherapy.

Currently, definitive subtyping requires immunohistochemical (IHC) analysis of tissue, often obtained through surgery or invasive biopsy. While ultrasound (US) is a standard, non-invasive tool for breast screening, human radiologists struggle to distinguish these molecular subtypes by eye because the visual differences are incredibly subtle and highly dependent on the operator's skill. This paper aims to bridge that gap using deep learning to extract "invisible" features from ultrasound images to predict subtypes preoperatively.

Methods

The authors proposed a hybrid DCNN architecture that fuses visual and clinical data:

  1. Image Feature Extraction: They fine-tuned three established architectures—EfficientNet, DenseNet-121, and VGGNet-16—pre-trained on ImageNet to extract high-dimensional features from B-mode ultrasound images.
  2. Clinical Integration: Clinical variables (patient age, tumor size, and aspect ratio) were encoded through a separate neural network into a 512-dimensional vector.
  3. Fusion: The image features and clinical vectors were concatenated and passed through a softmax layer for classification.
  4. Comparison: To validate the AI's "intelligence," they built a baseline BI-RADS lexicon model based on standard morphological descriptors (shape, margin, echo pattern) manually classified by experienced radiologists.

Datasets

The study utilized two distinct datasets to ensure the model's robustness:

  • Main Cohort (Center 1): 1,012 patients with 2,284 US images collected over an 11-year period. This was used for training and internal testing (4:1 split).
  • External Testing Cohort (Center 2): 117 patients with 153 US images collected from a different hospital. This served as a critical "real-world" test to see if the model could generalize across different ultrasound equipment and patient populations.

Key Insights and Ideas

  • AI vs. Human Lexicons: The DCNN consistently outperformed the BI-RADS lexicon model. This suggests that deep learning captures complex, non-linear textures in ultrasound images that human-defined descriptors (like "spiculated" or "oval") simply cannot capture.
  • The "Age and Size" Sweet Spot: The model isn't equally effective for everyone. It showed significantly higher accuracy in specific subgroups:
  • Older patients (>50 years): Better at identifying Luminal A subtypes.
  • Small tumors (≤20 mm): Particularly effective at distinguishing Luminal A from other types, which is historically difficult for radiologists because small nodules often lack distinct morphological features.
  • Triple-Negative Detection: The model showed promise in identifying Triple-Negative (TN) breast cancer—the most aggressive subtype—in younger patients (≤50 years), a group where early and accurate diagnosis is most critical.

Key Results

  • Luminal A vs. Non-Luminal A: In the main cohort, the DenseNet-121 model achieved an AUC of 0.717 and an accuracy of 80.1%, significantly higher than the BI-RADS model's AUC of 0.628.
  • Subgroup Performance:
    • For patients >50 years, the AUC for identifying Luminal A rose to 0.776.
  • For tumors ≤20 mm, the AUC reached its peak at 0.818.
  • External Validation: The model maintained a respectable AUC of 0.686 when identifying Luminal A in the external cohort (patients >50), proving it wasn't just "memorizing" the training hospital's specific imaging quirks.

Limitations and Future Work

The authors acknowledge several hurdles before this hits the clinic:

  • Retrospective Design: The data was collected after the fact; a prospective study is needed to see how the model performs in live clinical workflows.
  • Image Quality: Ultrasound is notoriously "noisy" compared to MRI. The model's performance is limited by the resolution and inconsistent standards of 2D ultrasound images.
  • Data Imbalance: Some subtypes (like HER2+) had much smaller sample sizes, leading to less stable predictions for those specific groups.
  • Future Directions: The team plans to explore multi-modal data, potentially adding color Doppler (blood flow) or elastography (tissue stiffness) to the input to provide the DCNN with a more holistic view of the tumor.

Summary updated

Abstract

We employed a DCNN model to predict the molecular subtypes of breast cancer based on US images. Our model can be valuable depending on the patient's age and nodule sizes.
BibTeX
@inproceedings{li2022preoperative,
  author = {Chunxiao Li and Haibo Huang and Ying Chen and Sihui Shao and Jing Chen and Rong Wu and Qi Zhang},
  title = {{Preoperative Non-Invasive Prediction of Breast Cancer Molecular Subtypes With a Deep Convolutional Neural Network on Ultrasound Images}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.848790},
}
Figure