Segmentation and Registration
of Pulmonary Nodules in CT

Chest Computer Tomography (CT) is used to diagnose pulmonary metastasis of oncology patients and evaluate disease progression or regression during treatment. Lung cancer remains a leading cause of cancer death in the Unites States. Early detection and treatment of lung cancer can significantly improve the long term health of those inflicted with it. A large number of people undergoing screening for lung cancer have small, potentially cancerous nodules in their lungs. These nodules are followed over time to determine potential size changes and evaluate their growth rates. Computer assistance has been proposed to handle inter- and intra-observer variations and also to decrease error in measurements.

We are developing and testing methods for the automatic segmentation and registration of nodule in the lung. These methods are either automations of what physicians have reported by hand or adaptations of methods proposed by others. Figures 1 and 2 below compare the segmentation obtained by a radiologist to segmentation obtained by two automated techniques. They demonstrate that a single segmentation technique is unlikely to work well for all cases. This is especially true when one considers considers a broader range of nodules such as ground glass nodules and/or patients with other diseases that affect the lung.

Figure 2: Poor Automatic Segmentation. Right: Radiologist Segmented Nodule. Middle: 2D Gradient Based Segmentation. Top: 3D Gradient Based Segmentation

Figure 1: Good Automatic Segmentation. Right: Radiologist Segmented Nodule. Middle: 2D Gradient Based Segmentation. Top: 3D Gradient Based Segmentation

Verification of the accuracy of segmentation methods is difficult. Small nodules in humans are rarely resected so their exact volumes are not know. Comparing segmentation obtained by computer to those obtained by physicians has become standard. We also test our segmentation methods on a lung phantom in which we know the exact properties of the simulated nodules. This is a useful measure of precision. However, the lung phantom does not completely mimic the true complexity of a human lung. Figure 3 below shows the difference between the phantom and typical human lungs.

Figure 3: Left: Lung Phantom. Middle: Initial Scan of Patient with Nodules Marked. Right: Follow Up Scan of Patient with Nodules Marked.

Aside from the problem of consistent measurements, physicians face another problem when attempting to find potentially malignant nodules in the lung. CT scanners are increasingly able to capture higher resolutions images that produce large amounts of information for the physician to sift through. If cancer screening becomes common, physicians might not be able to spend adequate time reviewing each image in a CT scan. We propose to help physicians by providing likely matches in follow up scans of nodules they have already identified in earlier scans. We are working on a system that will provide such matching ability based on a registration of lung surfaces combined with information about the individual objects segmented in the lung. Figure 4 below illustrates this.

Figure 4: Finding a match in a follow up scan from indentified nodules in an initial scan.

Segmentation and Registration of Pulmonary Nodules in CT

Segmentation and Registration
of Pulmonary Nodules in CT