New Large-Scale CT Data Set May Assist Universal Lesion Detection
The 32,000 image-data set may enable deep learning algorithms that can identify multiple lesion types using a single computing framework.
Identification of tumors and other lesions on radiological images is one of the most labor-intensive and time-consuming tasks in medical image analysis. Obtaining an accurate diagnosis requires detection and characterization of multiple lesion types in a single image. Deep learning-based artificial intelligence algorithms have the potential to speed and standardize this process. However, development of an automated multicategory or “universal” detection framework has been hampered by the lack of a large-scale data set with multiple identified lesion types.
In July, the National Institutes of Health’s Clinical Center released a large data set of computed tomography (CT) images assembled from archived medical images spanning two decades. The data set, called DeepLesion, contains more than 32,000 anonymized CT images from 4,400 different patients, making it much larger than existing public medical image data sets.
The images have been annotated by NIH radiologists to label significant clinical findings, such as lesion locations and measurements. The lesions include lung nodules and lesions in liver, kidney, bone and other tissues.
Public release of the data set is an important step toward enabling the creation of a universal lesion detector, the authors say. As an early demonstration, they used DeepLesion to train an automated lesion detection algorithm. In initial trials, the detector achieved greater than 80 percent sensitivity with five false positives reported per image. Upon later review, many of the reported false positives were determined to be true lesions that were unlabeled in the data set images.
The authors suggest this type of automated detector could serve as an initial screening tool to highlight images for review by radiologists or other detection systems specialized for different lesion types.