Deep Learning May Detect Breast Cancer Earlier than Radiologists

A deep learning model may be able to detect breast cancer one to two years earlier than standard clinical methods, according to a study published in Nature Medicine.

The tool also demonstrated promising generalizability, performing well when tested across populations and clinical sites not involved in training the algorithm.

Breast cancer is a devastating disease, with high mortality rates around the world. Screening mammography is estimated to decrease breast cancer mortality by 20 to 40 percent. However, researchers noted that significant false positive and false negative rates, along with high interpretation costs, leave room to improve quality and access.

To address these limitations, researchers have explored using deep learning in mammography. But obtaining large amounts of annotated data remains a significant challenge for training deep learning models for this purpose. Research teams also often have difficulty ensuring generalization beyond the populations represented in the training dataset.

The team set out to develop a deep learning algorithm that would be able to accurately identify breast cancer in screening mammograms and perform well in different populations. Researchers trained the model in a series of stages, testing the deep learning tool on tasks that became increasingly more difficult – a method that imitates how humans typically learn.

The group compared the performance of the deep learning model to that of five full-time, breast-fellowship-trained radiologists reading the same screening mammograms. The results showed that the tool outperformed all five radiologists, with an average increase in sensitivity of 14 percent.

These findings suggest that the algorithm could help detect breast cancer one to two years earlier than standard detection in many cases.

“Our results point to the clinical utility of AI for mammography in facilitating earlier breast cancer detection, as well as an ability to develop AI with similar benefits for other medical imaging applications. We have developed an approach that mimics how humans often learn by progressively training the AI models on more difficult tasks,” said lead author Bill Lotter, PhD, CTO, and co-founder of DeepHealth.

“By leveraging prior information learned in each successive training stage, this strategy results in AI that detects cancer accurately while also relying less on highly-annotated data. Our approach and validation extend to 3D mammography, which is particularly important given its growing use and the significant challenges it presents for AI.”

To further test the generalizability of the model, researchers evaluated its performance on a dataset obtained from an urban Chinese hospital. A team assessed the algorithm locally at the hospital, with data never leaving the site.

“Testing generalization to this dataset is particularly meaningful given the low screening rates in China and the known (and potentially unknown) biological differences found in mammograms between Western and Asian populations. For instance, there is a greater proportion of dense breasts in Asian populations, which can increase the difficulty of mammogram interpretation,” researchers noted.

The group found that the deep learning model – which was primarily trained on data from Western populations – generalized well to a Chinese population, achieving an area under the curve of 0.971.

The team expects that their study could improve the application of deep learning to screening mammography.

“We have developed a deep learning approach that achieves state-of-the-art performance in classification of breast cancer in screening mammograms. These results show great promise towards earlier cancer detection and improved access to life-saving screening mammography using deep learning,” researchers concluded.

As AI, machine learning, and other analytics tools become more widespread in healthcare, researchers are increasingly looking for new methods to train algorithms and ensure they will be effective across different populations.

A team from the University of Iowa (UI) recently received a $1 million grant from the National Science Foundation to develop a machine learning platform to train algorithms with data from around the world.

The grant will help UI researchers develop a decentralized, asynchronous solution called ImagiQ, which relies on an ecosystem of machine learning models so that institutions can select models that work best for their populations. Organizations will be able to upload and share the models, not patient data, with each other.

“Traditional methods of machine learning require a centralized database where patient data can be directly accessed for training a machine learning model. Such methods are impacted by practical issues such as patient privacy, information security, data ownership, and the burden on hospitals which must create and maintain these centralized databases,” said Stephen Baek, assistant professor of industrial and systems engineering at UI.

“ImagiQ will further federated learning by decentralizing the model updates and eliminating the synchronous update cycle. We are going to create a whole ecosystem of machine learning models that will evolve and improve over time. High performing models will be selected by many institutions, while others are phased out, producing more reliable and trustworthy outputs.”