Deep learning model to diagnose breast cancer

AI diagnosis of breast cancer

Date: 11th March 2020

Cancer is the leading cause of death worldwide, accounting for 9.6 million deaths in 2018.  Breast cancer (together with lung) is the most common cancer, and is the second leading cause of cancer-related death in women. Accurate identification of axillary lymph-node (ALN) involvement in patients with breast cancer is crucial for prognosis and therapy decisions and now scientists have successfully applied AI to predict ALN status in patients with early-stage breast cancer.

If breast cancer spreads, the lymph nodes in the underarm (the ALNs) are the first place it is likely to go.  The first lymph node to drain a primary cancer is called the sentinel lymph-node (SLN) and this is the first node to be biopsied, however, if positive for cancer, further axillary nodes also need to be removed, to determine how far the cancer has spread.

SLN dissection is often recommended to predict ALN status.  Whilst axillary dissection disrupts more of the normal tissue under the arm, both SLN and ALN removal can affect arm function and cause oedema. Therefore, non-invasive alternative ways to predict ALN status would be an ideal preferable option – these approaches, however, are currently limited.

Ultrasound (US) has been widely used to preoperatively characterise breast lesions and determine ALN status. However, there is a question regarding the accuracy of ALN ultrasonography‐negative results.

Some studies have also suggested that the stiffness of breast cancer tissue can be a predictor of ALN status and two-dimensional (2D) shear wave elastography (SWE) is a new technology to measure tissue stiffness.

This technique integrates a B-mode image (two-dimensional ultrasound image) with a color-coded map which shows the distribution of shear wave velocity (SWV). However, there are suggestions that it might be insufficient to evaluate ALN status accurately when used alone.

Recently, however, a relatively new field of medicine has started to emerge, radiomics, which is able to extract quantitative features from radiographic medical images using data-characterisation algorithms and this has the potential to provide an alternative solution to the problem.

Now a team of scientists, led by Jianhua Zhou, from the Collaborative Innovation Center for Cancer Medicine, China, have applied deep learning radiomics (DLR) of conventional ultrasound and shear wave elastography for predicting ALN status in preoperative patients with early-stage breast cancer.

This study, conducted between January 2016 and April 2019, enrolled a total of 1,342 women with 1,342 breast lesions only those meeting certain clinical selection criteria were included for further analysis, which in total amounted to 584 women with 584 malignant breast lesions.  The included patients were examined by conventional US and SWE, and had complete clinical information taken as the study required.

  • Deep-learning image-recognition base models were compared, the best performing model was then selected (ResNet50) and used as an encoder to which key clinical information was layered upon (Resnet50+C).
  • The enrolled patients were randomly divided into the training cohort and an independent test cohort. The training cohort were then used to optimise the model parameters.
  • The Resnet50+C method showed significantly better diagnostic performances in distinguishing patients with a negative axilla (N0) and patients with any axillary metastasis (N+(≥1)) than any single method.
  • The method was favourable in discriminating between patients with low metastatic burden of axillary disease (N+(1–2)) and patients with heavy metastatic burden of axillary disease (N+(≥3)).
  • Resnet50+C showed significantly better diagnostic performances in predicting ALN metastasis than the routine axillary US evaluated by an experienced radiologist.
  • With a false-negative rate similar to SLN dissection, this clinical parameter combined DLR might have the potential to serve as a noninvasively imaging biomarker to replace SLND for patients with early-stage breast cancer.

AI in heathcare

Conclusions and future applications:

The team here have developed and validated a clinical parameter combined DLR method based on breast conventional US and SWE images for preoperative prediction of ALN status in patients with clinical T1 or T2 breast cancer.  This non-invasive method should be a real step forward in predicting the metastatic extent of ALN.

The inclusion of clinical information was a key factor in improving the models performance.  However, one limitation of the study was it was a single-centre study. Acquiring more evidence from multi-centres would be needed to validate this model before clinical application in the future.

We are currently seeing a new wave of AI-driven diagnostic tools.  They seem particularly suited to image analysis; several types of cancer can now be predicted using AI and just last week we saw deep learning for the identification of COVID-19 pneumonia on high resolution CT. With health services around the world stretched, AI offers a valuable, accurate, and timing saving method of helping human experts make better clinical decisions.

The work presented here, may also offers us an insight into the nature of metastases spread.  There is also very powerful work being developed using a deep learning algorithm that enables the automated detection of metastases at the level of single cancer cells in whole mice.  Whilst this is currently being used on a ‘fixed’ animal, it is hoped that it will lead to a greater understanding of how a primary tumour can metastasise, and the unique way each cancer does so.  The team are also using it to accelerate development of new cancer treatments.

Together, deep learning is beginning to reveal a picture of each cancer type and how the disease progresses with time and thus providing us opportunities for eliminating or halting progression.


Zheng, X., Z. Yao, Y. Huang, Y. Yu, Y. Wang, Y. Liu, R. Mao, F. Li, Y. Xiao, Y. Wang, Y. Hu, J. Yu and J. Zhou (2020). “Deep learning radiomics can predict axillary lymph node status in early-stage breast cancer.” Nature Communications 11(1): 1236.