Why do we study image processing?
Images play a large role in the presentation of physiological information. Not all image data, however, are ideal. Images can be corrupted by noise, exhibit blur or spatial warping, contain non-optimal intensity/color representations, or simply be too large (or too small) to be of practical or diagnostic value. The following pages contain an overview of the following topics relevant to biomedical image processing:
- Image creation: generation of digital images and their resulting properties.
- Image modification: operations that can be performed on digital images including denoising, enhancement, and compression.
- Image analysis: techniques for extraction of biomedical data from images, including feature recognition.
The issues and techniques discussed here are general: they apply to images acquired using various methodologies, including X-ray radiography, X-ray computed tomography, MRI, ultrasound, and PET. Note that while image processing can be performed on analog images (e.g., by way of optical Fourier techniques), this overview concentrates on the processing of digital biomedical images.
Digital Image Creation
Analog images, having continuously varying color or gray-scale representations, can be stored as digital images for use in computer-based systems. The rules that apply to the digitization of one-dimensional (1D) signals (e.g., voltage versus time) also apply to the digitization of spatial images (e.g., gray level versus position): the images must be sampled and quantized with enough fidelity that they truly represent their analog image counterparts. This, of course, depends on the application for which the images will be utilized. Scanning, the initial step in the image digitization process, includes the division of the picture into a number of small regions called picture elements, or pixels (see Figure 1).
The scanning operation represents the image by a grid consisting of m rows × n columns. For example, in the image shown in Figure 1, m = 288 and n = 432, for a total of 124,416 pixels. Each pixel is addressed by its (m,n) coordinate within the matrix. The scanning operation is accompanied by a sampling operation, which transforms the representative brightness of each pixel into an analog voltage level or other measurable signal. In the case of an X-ray radiograph, this operation might be performed by a photomultiplier tube.
A good example of a common line scan sensor is a flatbed scanner, which uses an array of discrete silicon imaging elements, called photosites, to produce voltage output signals that are proportional to the intensity of the incoming light at each linear location. These solid state devices can be shuttered at high speeds (e.g., 1/10,000 of a second), and their precision typically ranges from 256 to 4096 elements [Gonzalez]. Solid state area sensors also exist, containing grid sizes from 32 × 32 pixels up to 1280 × 1024 pixels.
Finally, each pixel in the grid is assigned an integer that corresponds to its brightness level. This quantization operation is similar to the quantization performed on 1D signal data. For example, in an 8-bit gray-scale image, the integer 0 may represent black, the integer 255 may represent white, and the integers in between will represent various gray scales that correspond to image intensity. The general convention in gray-scale images is that larger integers represent brighter pixel values (or pixels exhibiting greater signal
strength).
Figure 1. Eight-bit, gray-scale digital image consisting of individual square pixels, or picture elements.
As illustrated in Figure 2, two-dimensional (2D) spatial images must obey sampling constraints that are similar to those imposed on 1D, time-based signals. When 1D signals are digitized with an analog-to-digital (A/D) converter, they must be sampled at a frequency greater than or equal to twice the highest frequency component in the signal. Likewise, 2D images must be sampled at a spatial frequency greater than or equal to twice the highest spatial frequency component in the image. Otherwise, these undersampled images will display the type of aliasing depicted in Figure 2. Note that to redisplay a digitally stored image, a D/A conversion is often required (e.g., for a cathode ray tube in a television set).
Figure 2. Simple illustration of aliasing in a digital image line [after Seeram, p. 55].
Some images are not produced on native rectangular grids. B-mode ultrasound images, for example, are often reconstructed from fan-shaped beams or radially-distributed ultrasonic reflectance data (see Figure 3). Production of a 2D image requires that these radial line data be interpolated. These interpolation schemes can affect the quality of the overall images, especially if the individual line scans are separated by large angles.
Figure 3. Examples of rectangular images constructed from non-rectangular imaging
modalities [Fetal Image: http://www.biosound.com/image5.html; OCT/IVUS: CLEO
’96 Proceedings, p. 55]
Why Digitize an Image?
The reasons for digitizing 2D image data are very similar to the reasons for digitizing 1D data, revolving primarily around (1) the ability to transmit image data “noise free” and (2) the potential for processing these images with computational tools. These tools
provide assistance in the following areas:
- Image Enhancement/Restoration: image improvement, possibly through the
reduction or removal of noise, artifact, or unnecessarily fine detail (e.g.,
processing (1) degraded images of unrecoverable objects or (2) images from
experiments that are too expensive to duplicate) - Image Analysis: extraction of information with/without interpretation
- Pattern Recognition: structures and patterns are “seen” and recognized
- Image Detection: identification of certain shapes, contours, or textures while
disregarding other image features - Geometric Transformation: rotation and/or scaling of the image
- Data Compression: Reduction in image size for storage or transmission
Part II of Biomedical Image Processing will discuss image properties and DICOM Image Standards. This information was compiled by Steve Warren of Kansas State University using the following references:
[1] Seeram, Euclid. Computed Tomography: Physical Principles, Clinical
Applications, and Quality Control, W.B. Saunders, Philadelphia, © 1994, ISBN 0-
7216-6710-4
[2] Shung, K. Kirk, Michael B. Smith, and Benjamin Tsui. Principles of Medical
Imaging, Academic Press, San Diego, © 1992, ISBN 0-12-640970-6
[3] Rosenfeld, Azriel and Avinish C. Kak. Digital Picture Processing, Second Edition,
Volume 1, Academic Press, San Diego, ©1982, ISBN0-12-597301-2.
[4] Rosenfeld, Azriel and Avinish C. Kak. Digital Picture Processing, Second Edition,
Volume 2, Academic Press, San Diego, ©1982, ISBN 0-12-597301-2.
[5] Gonzalez, Rafael C. and Richard E. Woods. Digital Image Processing, Addison-
Wesley, Reading, MA, ©1993, ISBN 0-201-50803-6.
[6] Teuber, Jan. Digital Image Processing, Prentice Hall, New York, ©1989, ISBN 0-
13-213364-4.