Vladimir S. Ostojic

Detection of collimation field in digital radiography using Frobenius norm of Hessian

Collimators are used to focus the radiation beam to a certain anatomical region in order to reduce the patient dose. Redundant shielded region should be removed before diagnostic image processing, as it can impair image enhancement algorithms. Manual selection of irradiated image region is impractical and should be automated. In this paper we propose a Frobenius norm of Hessian based algorithm for automatic collimation field detection. Algorithm validation was performed through, among others, maximum absolute difference between manually annotated and automatically determined collimation region. Our algorithm was tested on a database that contains 680 images and the percentage of correctly annotated regions is 96.03 %.

Low-value Laplacian pyramid coefficient prediction for faster radiography image processing

We propose a method for predicting low-value Laplacian pyramid details that can be used for radiographic image processing time reduction via reorganization of the Laplacian pyramid calculation process. High resolution low-value detail prediction is based on the value of coarser details. Low-value details are then omitted in Laplacian pyramid calculation as well as in the image enhancement process. Proposed algorithm was evaluated on a database consisting of real medical images. Image degradation introduced by the proposed method was measured using structural similarity index (SSIM). It was shown that average processing time reduction of 4.62 % can be achieved without noticeable processed image degradation. The average SSIM obtained on the database was 0.995.

Automatic evaluation of collimation field rotation angle in digital radiographic images

In this paper we propose a method for automatic collimation border rotation angle detection. Algorithm utilizes pooling of image gradients based on their orientations to form a histogram of oriented gradients (HOG) with the goal of determining the dominant gradient orientation in the image as collimation border rotation angle. To avoid accumulation of lower magnitude gradients only a percentage of the highest magnitude gradients was used during HOG calculation. Algorithm was evaluated on image database consisting of 244 real clinical images. Performance was measured through mean absolute error (MAE) between manually determined collimation field rotation angle and the angle estimated by the algorithm. Algorithm has achieved MAE of 0.388 degrees when using 21 % of the highest magnitude gradients.

An open-source digital diagnostic radiography image annotation software

As radiographic imaging is one of the most used medical imaging modalities, it has been a subject of various research in the domains of image processing and computer vision. Expert annotated databases are needed to enable easier algorithm development, validation and verification. From algorithm development standpoint, image information such as the existence and position of collimation, existence of directly irradiated foreground, image orientation and anatomy category can be useful. To the best of our knowledge, a database that contains all of that information does not exist, which is why we have designed a software for manual annotation of digital diagnostic x-ray images. The primary aim of this software is to aid researchers in creating expert annotated databases. The software is open-source and publicly available on-line.

Objective comparison of digital flat-panel radiography detectors

Digital flat-panel radiography detectors have become a standard part of modern digital radiography systems. Performance of a detector can be evaluated through three major indicators: noise, contrast and spatial resolution. Signal-to-noise ratio was employed to evaluate noise presence, contrast-to-noise ratio was utilized for contrast characteristic estimate while modulation transfer function and visual inspection of resolution phantom image were used for spatial resolution evaluation. In this paper, we present an extensive comparison of eight different flat-panel detectors using aforementioned evaluation methods. Comparison results are presented both numerically and graphically.

Automatic radiography image orientation using machine learning

Mobile digital radiography receptors, known as flat panels, apart from numerous advantages create an issue of proper image orientation. Common orientation of anatomical structures in radiography images is vital in reducing pre-diagnostic processing times. Various features and machine learning methods for determining current orientation of an image are examined with the aim of determining appropriate rotation of radiographic hand images. Obtained results are analyzed and further research directions are proposed.