Deep learning for real-time detection of breast cancer presenting pathological nipple discharge by ductoscopy

Feng Xu, Chuang Zhu, Zhihao Wang, Lei Zhang, Haifeng Gao, Zhenhai Ma, Yue Gao, Yang Guo, Xuewen Li, Yunzhao Luo, Mengxin Li, Guangqian Shen, He Liu, Yanshuang Li, Chao Zhang, Jianxiu Cui, Jie Li, Hongchuan Jiang, Jun LiuFrontiers in Oncology 2023 · DOI

TL;DR

Researchers have developed the Intelligent Ductoscopy for Breast Cancer Diagnostic System (IDBCS), a deep learning tool that identifies breast cancer during live ductoscopy exams. By leveraging a DenseNet-based architecture, the system achieved over 91% diagnostic accuracy, significantly outperforming human endoscopists and providing a real-time "second opinion" for clinicians.

What Problem Does This Paper Address?

Pathological nipple discharge (PND) is a frequent red flag for breast cancer, yet diagnosing the underlying cause is notoriously difficult. Ductoscopy—inserting a micro-endoscope into the breast ducts—allows for direct visualization of lesions, but the learning curve is steep.

Early-stage cancers often lack distinct endoscopic features, leading to high rates of missed diagnoses. Furthermore, there is a global shortage of expert endoscopists capable of interpreting these complex, low-resolution visual feeds. The authors sought to bridge this gap by creating an AI system capable of analyzing ductoscopy images in real-time to assist clinicians in identifying malignant tumors.

Methods

The authors developed IDBCS, a two-stage diagnostic framework:

  1. Image-Level Analysis: They tested four popular deep learning backbones (VGG, Inception-v3, ResNet, and DenseNet). DenseNet was selected as the winner due to its superior performance in feature extraction and inference speed.
  2. Patient-Level Diagnosis: Since a single ductoscopy procedure generates multiple images, the system uses a voting mechanism. It averages the probability scores of all images from a patient; if the average exceeds a specific threshold, the patient is flagged as having a high risk of malignancy.

To make the "black box" of AI more transparent, the team used Score-CAM (a class activation mapping algorithm) to generate heatmaps, highlighting exactly which visual regions the AI prioritized when making a diagnosis.

Datasets

This was a robust, multicenter study involving six hospitals in China.

  • Total Data: 11,305 ductoscopy images from 1,072 patients.
  • Training & Internal Validation: Data from Beijing Chao-Yang Hospital (BCYH) was split 8:2 for model development.
  • Prospective Validation: A separate set of 511 images from BCYH was used to test the model in a "real-world" forward-looking scenario.
  • External Validation: To ensure the model wasn't just memorizing one hospital's equipment quirks, it was tested on datasets from five other primary care hospitals (HDH, DLH, HRH, PGH, and YQH).

Key Insights and Ideas

  • AI vs. Human: The most striking insight is the performance gap. IDBCS achieved an accuracy of 91.2%, while even "expert" endoscopists (those with >5 years of experience and >500 exams) only reached 72.6%. This suggests that deep learning can pick up on subtle textural patterns invisible to the human eye.
  • The "Hemorrhage" Marker: Using Score-CAM, the researchers found that the AI frequently focused on areas of hemorrhage (bleeding). They discovered a high correlation (mIoU of 0.598) between malignant tumors and localized bleeding, confirming that the AI learned a clinically relevant biological marker.
  • Real-Time Feasibility: The system processed roughly 31 images per second (32ms per image) on an Nvidia Tesla T4 GPU. This is crucial because it allows the AI to provide feedback during the procedure, rather than as a post-hoc report.

Key Results

  • High Diagnostic Power: The Area Under the Curve (AUC) remained consistently high across all tests: 0.975 for internal validation and ranging from 0.922 to 0.965 across the five external hospital sites.
  • Superior Sensitivity: In the prospective group, IDBCS showed a sensitivity of 84.38%, compared to just 25.00% for expert endoscopists. This means the AI is far less likely to miss a cancer that is actually present.
  • Robustness: The model maintained high specificity (>80% in all datasets), ensuring that it doesn't just "flag everything" but remains precise in its detections.

Limitations and Future Work

While the results are promising, the authors noted a few hurdles:

  • Subtype Differentiation: While IDBCS is great at distinguishing "benign vs. malignant," it struggled to differentiate between specific cancer subtypes, such as Ductal Carcinoma In Situ (DCIS) and Invasive Breast Cancer (IBC), likely due to a smaller sample size for those specific categories.
  • Demographic Bias: The training data was primarily from Northern Chinese populations. Further validation is needed to ensure the model performs equally well across different ethnicities and geographic regions.
  • Clinical Integration: Currently, the model only looks at images. Future versions could improve accuracy by integrating clinical data like patient age, duration of symptoms, and the color of the nipple discharge.

Summary updated

Abstract

IDBCS outperforms clinical oncologists, achieving high accuracy in diagnosing breast cancer with PND. The novel system could help endoscopists improve their diagnostic efficacy in breast cancer diagnosis.
BibTeX
@inproceedings{xu2023deep,
  author = {Feng Xu and Chuang Zhu and Zhihao Wang and Lei Zhang and Haifeng Gao and Zhenhai Ma and Yue Gao and Yang Guo and Xuewen Li and Yunzhao Luo and Mengxin Li and Guangqian Shen and He Liu and Yanshuang Li and Chao Zhang and Jianxiu Cui and Jie Li and Hongchuan Jiang and Jun Liu},
  title = {{Deep learning for real-time detection of breast cancer presenting pathological nipple discharge by ductoscopy}},
  booktitle = {Frontiers in Oncology},
  year = {2023},
  doi = {10.3389/fonc.2023.1103145},
}
Figure