Matthew Freedman, MD, MBA

Computer-aided detection (CAD) for chest radiographs is a method for assisting radiologists in their interpretation of these studies. Currently, the method focuses on assisting radiologists to detect small lung nodules that could represent small lung cancers. The commercially available system was validated for FDA premarket approval in a study using primary non-small cell lung cancer (NSCLC). Another product (not available in the United States) is designed to assist radiologists detect small areas of pneumonia in patients with symptoms of severe acute respiratory syndrome (SARS).

Computer-aided diagnosis (CADx) methods are devised to assist radiologists to determine the disease that has caused an abnormality; for example, helping to determine whether a lung nodule identified on a chest radiograph is due to cancer or is benign. These systems are under development, but are not yet commercially available.

Figure 1. The computer system has made circles around areas of the chest radiograph. The upper circle marks an area of rib and clavicle overlap and is a false positive. The lower circle identifies the location of a 17 mm lung nodule. (Click the image for a larger version.)

CAD systems use a computer algorithm to analyze images for signs of specific diseases. For the computer algorithm to analyze the image, the image must be in digital form. Converting an analog film into digital form is done with a film scanner (digitizer). Chest radiographs obtained with computed radiography and digital radiography systems are already in digital form and can be processed directly. In the commercially available system, more than 80 mathematical features are used to identify suspect areas and to eliminate areas that are similar to lung nodules, but are due to other types of patterns on the image. Figures 1 and 2 show typical outputs of the system with marks on a small primary lung cancer and a small stellate nodule or scar.

Figure 2. A computer-generated circle surrounds a stellate nodule or nodular scar high in the right lung apex. Elsewhere on this image there were two false-positive circles. (Click the image for a larger version.)

All existing commercial CAD systems for lung and breast imaging involve tradeoffs in design. All of them can correctly identify the location of cancers. In this case the mark is called a true positive. All of them fail to identify the location of some cancers. These are called false negatives. All of them mark some areas as being of concern when no disease is present. These location marks are called false positives.

All existing commercial CAD systems have been designed to be used as a second reader, augmenting the radiologist’s primary detections by marking areas for repeat inspection by the radiologist. When the radiologist identifies a suspect location only after using the computer algorithm’s output, then the radiologist has benefited from the use of the system.

There have been two studies of the commercially available system that has been reviewed and accepted by the FDA.

In the first study, chest radiographs obtained in the NCI-sponsored Johns Hopkins Early Lung Cancer Program were sorted, checked for quality and then reevaluated by 15 radiologists. There were 80 primary non-small cell lung cancers ranging in size from 9 to 27 mm randomly intermixed with 160 chest radiographs from the same smoking population that had been shown by long-term follow-up to be cancer-free. This study included only males. Machine sensitivity was 66% with an average of 5.3 false positives per image; 15 radiologists interpreted the same set of cases without and with CAD and showed initial (without CAD) detection sensitivity ranging from 57.50% to 77.50% with a range of improvement when using the CAD prompts of 2% to +15%. In the clinical trial for the FDA-approved chest radiograph CAD system, 10 of 15 radiologists improved their detection of cancer with the system, four showed no improvement, and one decreased his detection of cancer by one case by selecting another nodule on the chest radiograph that was benign.

The second study used contemporary digital chest radiographs and included both men and women; 79 cases contained solitary malignant nodules of both primary lung cancer and solitary metastases. For all cases, the machine sensitivity was 63% with 5.0 false positives per image. For the primary NSCLC cases, the sensitivity was 67% with 5.0 false positives per image and showed equal sensitivity for lung cancer in men and women.

The system comes in two major configurations, one to be used with screen-film chest radiographs; it includes a film digitizer. The other is for chest radiographs that are already in digital form. The system can be freestanding and provide the results of the CAD analysis on a paper print, it can be connected to a single workstation, or it can be connected to a PACS. In the clinical trial for FDA approval, the radiologists spent an average increase of 8 seconds when reading with CAD. This included the time needed to enter their interpretations into the computer.

A review of recent advances in chest radiography, including CAD, is available. 1

Matthew Freedman, MD, MBA, is a radiologist and associate professor of oncology, Georgetown University Medical Center, Washington, DC. He started his research in computer-aided detection of breast and lung cancer in 1991 and has many publications on this topic. He is a consultant to Riverain Medical Group and was clinical director for the clinical trial submitted to the US Food and Drug Administration for premarket approval, which was given.

References:

  1. Freedman MT. State-of-the-art screening for lung cancer (part 1): the chest radiograph. Thoracic Surgery Clinics. 2004;14:43-52.