top of page
  • Writer's picturePre-Collegiate Global Health Review

Artificial Intelligence in the Detection of Breast Cancer

By Aditya Syam, Jamnabai Narsee International School, Mumbai, Maharashtra, India

Breast cancer is the most frequently occurring type of invasive cancer (i.e., it spreads beyond the tissues in which it develops) and is also the second leading cause of cancer death in women. In 2018, 2.1 million women were estimated to have been diagnosed with breast cancer worldwide, and 630,000 of them died as a result (“AI beats doctors in detecting breast cancer”, 2020).

One of the major hindrances in the treatment of breast cancer is the diagnosis of the cancer at a late stage, beyond which effective treatment becomes more and more improbable. In many cases, this is because the patient’s mammogram (a method of breast cancer screening and diagnosis) may seem normal despite the presence of a mass of cells. According to a study, 36% of patients that experienced delayed diagnosis had normal mammograms as compared to only 7% of patients who were diagnosed early (Tartter et al., 1999).

The advent of deep learning and AI-based technologies has been harnessed in multiple projects to test the validity of these methods in breast cancer screening, to obtain more precise results and reduce the number of false positives.

When radiologists read mammograms, they usually look for specific kinds of lesions (a form of calcification cluster or other tissue findings) that can indicate the presence of a malignancy. The established method of computer-aided detection relies on human intervention and programming to determine which parts of the image may signal a lesion (Sechopoulos et al., 2020).

On the other hand, in Artificial Intelligence, convolutional neural networks that use deep learning do not rely on such intervention and essentially teach themselves what signifies a malignant lesion in a mammogram. Each mammogram image is repeatedly simplified based on its spatial features to identify possible lesion areas. The algorithm would previously have been trained using innumerable sample mammograms, based on which its internal parameters would be modified to decide what does or does not count as a malignant lesion (Sechopoulos et al., 2020).

A joint study by Imperial College London and Google Health used a computer algorithm which was able to outperform six radiologists in accurately scanning and reading mammograms. The model had been trained using deep learning methods with sample mammograms from 29,000 women (“AI beats doctors in detecting breast cancer”, 2020). The AI model reduced the number of screening and diagnosis errors – false positives were reduced by 1.2% and false negatives by 2.7%. The accuracy levels equaled or exceeded that of two doctors separately evaluating the same mammogram (O'Hare, 2020).

A South Korean company, Lunit, that uses AI in healthcare applications, developed an algorithm called Lunit INSIGHT MMG, that had been trained by 72,000 previous breast cancer mammograms (“Recent Studies Reveal High Performance of Lunit AI in Breast Cancer Detection”, 2020). It was created with a special focus in mind – to divert radiologist’s attention and resources towards mammograms which show the possibility of a cancer and away from the mammograms which are clearly negative for the presence of lesions. This way, the radiologist workload could be reduced, and detection can be improved (“Recent Studies Reveal High Performance of Lunit AI in Breast Cancer Detection”, 2020).

A study performed by the Karolinksa Institute from Sweden concluded that the Lunit AI showed much higher levels of accuracy than others of its kind (15% higher sensitivity, i.e., ability to detect true positives). Further, it showed real potential for use in a clinical setting – the AI performed better than radiologists and achieved a combined high sensitivity rate of 88.6% when used together with one radiologist reader (“Recent Studies Reveal High Performance of Lunit AI in Breast Cancer Detection”, 2020).

The study also revealed that if radiologist resources were not employed for the lowest 60% of AI scores (that indicate minimal possibility of cancer) but instead diverted to the top 2% of cases with higher possibilities, then approximately 103 of 547 cases in the study would have been detected 2 years earlier (“Recent Studies Reveal High Performance of Lunit AI in Breast Cancer Detection”, 2020).

Now, in an even more advanced step, MIT’s Computer Science and Artificial Intelligence Laboratory has developed an image-based AI model that can predict breast cancer up to 5 years in advance. The model was trained to identify minute patterns in the breast tissues that tend to be precursors to the development of cancer. This approach can also be used to personalise periodic screening for women depending on the risk of cancer development – the model was able to accurately classify 31% of cancer patients in the highest risk category, compared to an average of 18% for other models (Conner-Simons & Gordon, 2019).

MIT’s model is also aimed at making cancer detection more equitable – most earlier models were trained using data from only white populations, and so were less accurate for patients from other racial backgrounds. Further, black women are said to be 42% more likely to die from breast cancer due to lack of access to cancer healthcare. The current model, however, works equally well for people of different races (Conner-Simons & Gordon, 2019).

Although the use of AI in breast cancer detection has not been normalised yet and has been used primarily in experimental studies and pilot projects, the higher accuracy rates showcased by the AI models indicate that it has potential to be introduced in the clinical setting. Of course, multiple barriers will have to be cleared before this AI is commercialised and introduced. These include clinical barriers like patient acceptability and the training that radiologists will need to be able to incorporate AI into their clinical practices. The risk of overdependence on AI would be ever-present and the automated output might risk creating a pre-conceived bias for the radiologist (Hickman et al., 2021).

On the other hand, there are also ethical barriers such as the storage and transfer of patients’ data in case the algorithm is being run from an offsite system (Hickman et al., 2021). However, when AI is truly incorporated into the clinical setting, it could significantly reduce the amount of healthcare resources expended on cancer detection and improve the rates of breast cancer survival.



Centers for Disease Control and Prevention. (2020, September 14). What Is Breast Cancer? Centers for Disease Control and Prevention.

Conner-Simons , A., & Gordon, R. (2019, May 7). Using AI to predict breast cancer and personalize care. MIT News | Massachusetts Institute of Technology.

Felman, A. (2019, August 12). Breast cancer: Symptoms, causes, and treatment. Medical News Today.

Hickman, S. E., Baxter, G. C., & Gilbert, F. J. (2021, March 26). Adoption of artificial intelligence in breast imaging: evaluation, ethical constraints and limitations. Nature News.

ITN. (2020, December 2). Recent Studies Reveal High Performance of Lunit AI in Breast Cancer Detection. Imaging Technology News.

Knappily. (2020, January 2). AI beats doctors in detecting breast cancer. Knappily.

McKinney, S. M., Sieniek, M., Godbole, V., Godwin, J., Antropova, N., Ashrafian, H., … Shetty, S. (2020, January 1). International evaluation of an AI system for breast cancer screening. Nature News.

O'Hare, R. (2020, January 1). Artificial intelligence could help to spot breast cancer: Imperial News: Imperial College London. Imperial News.

Sechopoulos, I., Teuwen, J., & Mann, R. (2020, June 9). Artificial intelligence for breast cancer detection in mammography and digital breast tomosynthesis: State of the art. Seminars in Cancer Biology.,mammography%20and%20digital%20breast%20tomosynthesis.

Tartter, P. I., Pace, D., Frost, M., & Bernstein, J. L. (1999, January). Delay in diagnosis of breast cancer. Annals of surgery.

Article Thumbnail:

"Breast Cancer Cells" by National Institutes of Health (NIH) is marked under CC PDM 1.0. To view the terms, visit


bottom of page