Automatic Collimation in Peripheral X-Ray Imaging

A method has been developed using which the foreground (body) in a fluoroscopy image of the peripherals can be segmented from the background (existing collimation and direct exposure). The extracted information about where the body is in an image is used to suggest settings for the collimator hardware during image acquisition. Segmentation is achieved through global feature extraction and supervised classification. 1 Background of the application Preventing direct exposure and scattered x-rays from reaching the image intensifier or the image recording media is a major concern in x-ray diagnostic imaging. An important measure employed for this is collimation of the beam, using x-ray opaque materials, to that minimally necessary for imaging the object in question (see Figure 1). This eliminates scattered radiation from irrelevant structures and is an important measure from a radiation hygiene point of view. Moreover, without collimation, the direct exposure of x-rays to the image intensifier could damage the device and produce diagnostically useless images. A typical current peripheral angiography study of the legs requires three full scans from abdominal to distal foot region [3]. Each scan consists of taking images at several stations (positions) along the leg. First, the fluoroscopic control scan allows the physician to check individual stations, measure patient transparency and manually set the collimator. As the physician sets the collimator at each station to optimally cover the non-body regions of the image, the settings are saved into a computer lookup table. The second scan, precontrast scan, produces reference (mask) images, for example to be used in digital subtraction angiography (DSA). During the final scan, contrast scan, a bolus is injected into the patient and the imaging equipment follows the flow of the bolus taking images at each station. During this image acquisition, the collimator settings are retrieved from the lookup table and applied at each station. Automating the process of setting the collimation saves time during examination. It also reduces the effect of human factors such as subjectivity and fatigue. While the physician is involved in the important activity of caring for the patient, it is preferable to have as few tedious distractions as possible. 2 Image processing problem To perform automatic collimation, it is important to find reliably, precisely and efficiently where the patient’s body is in an image. If the precise location of the body is known, the collimator can be adjusted to cover as much of the nonbody region and as little of the body region as possible, given the hardware constraints. A few researchers have considered the problem of segmenting the foreground and background in computed radiography x-ray images. Zhang and Huang [7] address the problem of collimation recognition and removal, to improve the display of x-ray images in a Picture Archival and Communication System (PACS). Their method, based on statistical models of collimated regions and edges, is highly accurate. Researchers at Eastman Kodak have worked on the problem of skin tone scale improvement. Two sub-problems they consider are collimation detection [4] and skin line detection [6], the later being related to finding direct exposure regions. Kodak’s algorithms for skin tone adjustment have been clinically validated. The above works deal with computed radiography images, which are very different from fluoroscopy images. The task of segmenting the body in an x-ray fluoroscopy image at acquisition time is difficult for several reasons: Segmentation needs to take place without knowing which part of the body is being looked at. For some parts of the body such as the pelvis, the image may not contain any soft tissue boundaries or direct exposure regions. There may be images with more than one body part (e.g., legs) in it. The shape of the collimation and background can vary. It is desirable to have a single segmentation method that can handle all these cases. Low radiation dosage in a fluoroscopy study implies low signal to noise ratio of the images and poor contrast. Figure 1: A simplistic view of x-ray image acquisition Soft tissue boundaries often have very subtle intensity contrast and different “signatures” than step edges. Local intensity characteristics at a pixel are often inadequate for determining if it belongs to the body or the background, due to factors like soft (semitransparent) collimation and noise. For example, see the area between the legs in the last station in Fig. 3. Segmentation and automatic collimation need to be done at image acquisition time, placing tight constraints on the complexity and speed of the image processing operators that can be used.