Historically, researchers interested in developing automated methods to detect tumors in medical images have been hampered by the small size of the image collections available for training models using a deep-learning approach. Recently, however, NIH researchers unveiled DeepLesion, a free collection of over 30,000 annotated CT scans that scientists can use for this purpose.

Researchers have long wanted to develop an automated method to detect tumors in medical images. However, gathering enough scans to train a tumor-detection model to accurately recognize lesions using a deep-learning approach has been challenging. Recently, Ke Yan, PhD, a postdoctoral fellow at the NIH, and colleagues compiled DeepLesion, a dataset to address this problem (see nihcc.app.box.com/v/deeplesion). The dataset, which is available for free online, currently consists of 32,120 annotated CT scans featuring 32,735 cancerous and noncancerous lesions of various types, collected from 4,427 unique patients.

In addition to being the largest such dataset available, DeepLesion is also the broadest. “It includes a wide variety of lesion types and locations,” says Ronald Summers, MD, PhD, chief of the Imaging Biomarkers and Computer-Aided Diagnosis Laboratory at the NIH Clinical Center and senior author of the paper describing DeepLesion (J Med Imaging 2018;5:036501).

“Many of the existing datasets that have been made public are tailored to specific types of lesions,” explains Summers. Thus, the resulting tumor detection models target only one type of lesion, such as lung nodules or colonic polyps. With a broader training set, researchers may be able to develop a single model capable of identifying many tumor types.

Furthermore, tumors pictured in the DeepLesion scans have been measured by radiologists in accordance with RECIST guidelines. This type of high-quality annotation also sets DeepLesion apart from existing imaging datasets, says Jean-Emmanuel Bibault, MD, of Université Paris Descartes in France. “A major problem with deep learning in medical imaging has been the low quality of image annotation in most image databases. High-quality annotations are crucial for training models.”

graphic

A CT scan included in the DeepLesion database shows multiple liver tumors. Red and blue lines bounded by the green box measure one of the tumors using RECIST criteria. An automated method for detecting and measuring lesions could someday determine whether a patient's cancer treatment is working.

Summers says his lab will continue adding scans to the dataset and improving software to detect lesions and measure their diameters. However, as the scientific community works to improve tumor-detection models, Summers and Bibault say it will be important to start thinking about how automated tumor detection could be integrated into clinical practice.

“It is not clear how we will clinically validate promising algorithms,” notes Bibault. “Should we use randomized clinical trials? If so, should the models be ‘frozen,’ meaning we don't add any more data to train them? Or should they be continuously updated, so we don't lose one of the advantages of deep learning?”

“We must also consider how automated detection could be implemented in our medical system in a helpful and cost-effective manner,” says Summers.

Bibault agrees. “What is the economic model for reimbursement for this kind of algorithm?” he asks. “We simply don't know how it should be done.” –Kristin Harper

For more news on cancer research, visit Cancer Discovery online at http://cancerdiscovery.aacrjournals.org/CDNews.

©2018 American Association for Cancer Research.
2018
American Association for Cancer Research.