Image blur, introduced by patient motion, is one of the most frequently cited reasons for image rejection in radiographic examinations. This problem becomes even worse for images that require long exposure times, such as lateral chest, lumbar spine, and abdomen radiographs. The detection and severity assessment of motion blur can be challenging. This is because inconsistent evaluation of motion in radiographs exists among technologists – and radiologists, too – due to the lack of accepted standards and various backgrounds of education and training. In addition, technologists, when performing visual quality assurance (QA), generally view images on relatively low-resolution monitors that are used by remote operations panels (ROPs) in the exam room or other preferred location. These monitors provide limited image quality for assessing motion blur in radiographs, compared with radiologists’ diagnostic-quality displays. Consequently, slight to moderate motion blur may go undetected during the visual QA step, which can cause images with limited diagnostic quality to be released to picture archiving and communication systems (PACS), and affect image diagnosis.

The goal of this work is to develop an algorithm with the capability to automatically detect motion blur in digital radiographic images as a means to aid technologists when performing QA.

To achieve this goal, the proposed motion blur detection algorithm follows the manner in which radiologists interpret images, and consists of four steps. First, a radiographic image is reoriented according to a predetermined hanging protocol (or radiologists’ preferred orientation). Next, by using exam type-specific rules, the algorithm locates the primary anatomy in the radiograph and extracts the most indicative region for motion blur detection. For robustness, the detection of the region of interest (ROI) is designed to be invariant to the size and pose of patients. The third step computes a set of features from the extracted ROI. These features capture the motion-sensitive information of the anatomy. Finally, the extracted features are evaluated by a set of classifiers trained to detect image blur for different anatomical structures, such as bone and soft tissue. The output of a classifier is a figure of merit that is associated with the probability of motion blur.

The algorithm is developed using a database of CR images that was developed over a one-year period, based on sequential case collections at two high-volume imaging centers. The database was populated with both QA-accepted images (i.e., images determined by the radiographic technologist during the visual QA process as having acceptable diagnostic quality) and clinically rejected images (i.e., images deemed by the radiographic technologist as not suitable for diagnosis). Validation was achieved through visual review by a trained expert.

The performance of the algorithm is evaluated against a total of 550 images (300 with no visibly detectable motion and 250 images graded as having either moderate or significant motion blur). These test images cover a wide variety of exam types, especially those requiring long exposure times. Preliminary experiments show promising results with 82% detection sensitivity, 72% specificity, and an overall accuracy of 76%.

Determining motion blur in a radiograph is a difficult problem because the quality of radiographs varies considerably due to the wide variety of examination conditions. Thus, the key to solving this problem is to find the image features that are only sensitive to motion blur – regardless of variations in patients’ situations (e.g., size and pose), and differences in examination setting. The proposed method provides a solution to the problem by adaptively detecting ROI in the radiographs and extracting motion blur sensitive features from ROI. The solution works effectively for most exam types.

Another challenge for image blur detection is related to the different sources of the motion blur themselves. Basically, there are two common sources of anatomical motion blur in radiographs. One blur source comes from patient movement during image acquisition, namely external motion. The result of external motion is a blurred appearance of the entire anatomical region in the image. The second source relates to internal motion that is due to the normal involuntary functioning of the anatomy. For example, the beating of the heart can cause some amount of blur either directly, if the heart tissue lies within the image, or indirectly, by the compensating for movement of surrounding structures. This effect can easily result in blur within local regions of a chest radiograph. The proposed method successfully addresses the problem by taking advantage of exam type information into ROI detection, and adaptively adjusting the shape of the ROI using the exam type-specific rules.

Motion blur proves to be more difficult to detect images from in-patient units, such as intensive care units (ICU) or emergency rooms (ER). This is because in-patient images are generally captured using portable devices, and tend to have more foreign artifacts, such as tubes and wires. In addition, the variability in normal anatomical structures is considerable compared to the differences in unsharpness that would distinguish a diagnostically acceptable image from one that should be rejected.

Given the diversity and variability we stressed here, and that only a preliminary feature set was used, the initial performance is encouraging. The number of false-positive detections is expected to decrease greatly as additional features are utilized, and as a greater number of images are added to the training database.