Lung cancer is the leading cause of cancer-related deaths in the U.S. resulting in more deaths than the number of deaths from colon, breast, and prostate cancers combined. Early detection of lung cancer allows for timely therapeutic intervention and more favorable prognoses for patients. Currently, computed tomography (CT) is the most accurate imaging modality for lung cancer screening, but in a cancer screening program, radiologists are likely to overlook some lung cancers because they have to read a large number of images.  A computer-aided detection scheme (CADe) automatically identifies potential lung nodules in CT and is thus particularly useful in reducing the chance of overlooking lung cancer.

We have developed innovative key techniques to automatically detect lung nodules in CT, including a nodule enhancement filter to selectively enhance nodules and suppress normal structures [Med. Phys. 30: 2040-2051, 2003], an accurate nodule segmentation method for feature calculation and nodule quantification [Med. Phys. 34: 4678-4689, 2007], and a novel rule-based classifier to achieve both good performance and low overtraining [Med. Phys. 33: 868-875, 2006]. Figure 1 shows three low-contrast lung nodules that were detected by our CADe scheme. By looking at the detected nodules in CT, radiologists can confidently diagnose all detected nodules and thus reduce the chance of missing a malignant nodule.

	(a)		(b)		(c)
	Fig.1. Three low-contrast lung nodules with ground-glass opacity that were automatically identified by our CADe.  Nodules are indicated by white crosses.

Once  a  lung  nodule  is  detected, the  next  step  is  to  diagnose  it  as  either  cancer or benign. Currently, integrated positron emission tomography and computed tomography (PET/CT) is the most effective imaging modality for lung cancer diagnosis by providing complementary metabolic information (from PET) and anatomic information (from CT). We developed three computer-aided diagnosis (CADx) schemes to diagnose lung nodules in PET alone, CT alone, and both PET and CT [J Nuclear Medicine 47: 1075-1080, 2006]. We found that our CADx based on both PET and CT was significantly better in diagnosing lung cancer than was the CADx based on either modality alone. Figure 2 shows (a) CT and (b) PET images of a lung cancer patient. Because the cancer had very low metabolic activity in the PET image, our PET CADx incorrectly assigned a low likelihood of malignancy (0.27) for the lung cancer, which indicated a false negative finding. However, in the CT image, the cancer was presented with clear lung cancer pattern called spiculation; When the CT information was combined with the PET information, the likelihood of malignancy was increased to 0.84 and indicated a correct positive finding for lung cancer.

	Figure 2. (a) CT and (b) PET scans of a 69-year-old woman with a lung cancer. Using only the PET data would have resulted in missing a cancer that required treatment. The probability that the detected nodule is malignant, using only the PET data, is 0.27, but using both PET and CT data the probability increases to 0.84.