Gray Level Display Requirements Based on Visual Perception of Radiographic Patterns - SIIM News Spring 2009

Today’s medical radiographs are commonly displayed on LCD monitors and are calibrated to produce gray levels for which the luminance versus display value (DV) follows the DICOM Grayscale Display Function (GSDF). (1) The majority of current systems use a set of 256 gray levels (8-bit) selected from a palette of 766 to 3064. For more gray levels, all stages of the display pipeline must support higher definition. Some mammography products provide specialized application, graphic card, and monitor support of 1024 gray levels (10-bit). For a calibration resulting in a maximum luminance of 500 cd/m2 and a luminance ratio of 350, there are 619 DICOM Just Noticeable Difference (JND) indices with 2.43 JNDs per DV. The contrast per DV (i.e. deltaL/L) ranges from 0.016 to 0.049. Since the contrast associated with a 1.0 JND indice change represents the ‘just noticeable’ contrast of a small sinusoidal pattern, concern has been raised as to whether an 8-bit grayscale is sufficient for interpreting medical radiographs. This concern is qualified by the fact that sinusoidal patterns of the type used in psychovisual reasearch are not typical of the patterns found in medical radiographs. (2)

The visibility of the quantum mottle patterns in a radiograph can be taken as an indication of whether the gray level precision is sufficiently small for interpretation. As a pattern that interferes with low contrast signals, acceptable levels of quantum mottle are established in relation to patient exposure. It follows that if the gray level precision is sufficient to detect the quantum mottle pattern, it will be sufficient to interpret the radiograph.

Using medical LCD monitors (0.207 mm pixel pitch, 2048 x 1536 array), we have studied the just noticeable contrast for quantum mottle patterns representative of those found in medical radiographs. Visual detection was measured using automated two alternative forced choice (2AFC) experiments that determined the observer’s contrast threshold from about 200 images presented in 12-15 minute tests. The 2AFC observer experiment provides an effective way to measure the contrast threshold of image patterns. Experiments have shown that the contrast threshold measured with this method is less than that found by the method of variable adjustment used in much of the early psycho-visual research. The results were analyzed after each test using maximum likelihood estimation (MLE) methods. This implementation provides immediate feedback of results and is found to help observer satisfaction.

For this work we used the fine grained noise pattern of a direct digital radiography (DR) system and the somewhat smoothed noise pattern of a computed radiography (CR) system as models for the patterns used in 2AFC tests (see Figure 1). Specialized calibration methods were used to achieve a grey level precision of less than 1.0 JNDs per DV. In order to measure the visibility of noise in relation to the JNDs per DV, the relative standard deviation of the pixel values in each target image was used as a “noise contrast” value, which could then be converted similar to the definition of a JND and compared to a 1.0 JND shift.

Figure 1: Gaussian distributed noise pattern similar to noise seen in direct radiography (left) and computed radiography (right).

For the detection of CR noise, the mean noise contrast threshold from 2AFC testing was 4.6 times the contrast of 1.0 JND (N=7). For DR noise, the corresponding ratio was 8.5. For DR noise, the high spatial frequencies in the pattern were above those that can be seen by the human visual system, thus the DR noise patterns were less visible compared to the smoothed CR noise patterns with the same standard deviations.

The quantum mottle contrast of a digital radiograph presented on a display monitor varies depending on the exposure, the image processing, and the grayscale rendition used. For reference, we have measured the quantum mottle contrast in the noisy (i.e. radio-opaque) regions of abdominal radiographs exposed, processed, and presented using the clinical protocols in use at Henry Ford Health System. For CR images, the mean quantum mottle contrast (1 standard deviation) is 10.5 (N=6) times the contrast of 1.0 JND and for CR the corresponding contrast ratio is 15.4 (N=6).

Our previous work has shown that 2AFC tests result in a contrast threshold for sinusoidal patterns that are about 2/3 of the contrast associated with a DICOM JND.(3) In comparison, the contrast threshold for quantum noise patterns is seen to be 5 times that of a DICOM JND. This suggests that requiring a gray scale precision of 1.0 JND is an overly stringent criteria. (4)

The use of 1 standard deviation to describe the contrast of a noise pattern is conservative in that much of the image scene extends to 2 standard deviations and the maximum and minimum values of the pattern extend to 3-4 standard deviations. Thus, quantum noise patterns are effectively presented with gray levels spaced at about half the contrast threshold measured with the 2AFC experiment. Using this as a requirement, current 8-bit gray level displays appear to be adequate for the interpretation of radiographs.

For the abdominal radiographs considered to date, the noise contrast is about twice that of the human visual contrast threshold for similar noise patterns. This provides further support for the use of 256 gray levels. However, musculoskeletal extremity and breast images typically have a quantum mottle pattern that is less visible than for abdomen images when processed and displayed with contemporary protocols. Further work is required to understand the requirements for these procedures.

In conclusion, based on requirements defined by the human visual contrast threshold of quantum mottle patterns, 8-bit grayscale systems appear to be adequate for general radiography interpretations. Systems with more gray levels and better gray level precision may be beneficial for more demanding modalities such as mammography and musculoskeletal radiography.

Dr. Flynn is a Sr. Staff Medical Physicist at Henry Ford Health Systems in Detroit and an Adjunct Professor at the University of Michigan. His prior work has involved medical image display performance, gray scale performance, and the human visual system. He is also actively involved in PACS, Digital Radiography, and Computed Tomography. Dr. Tchou is currently a diagnostic medical physics resident at MD Anderson Cancer Center in Houston.  He graduated from the University of Michigan in 2007 with a doctorate in Nuclear Engineering and Radiological Sciences.  His graduate research was done under Dr. Flynn at Henry Ford Health System in Detroit, where he worked on visual performance diagnostics for radiology.

Editor’s note: This article is based on a scientific poster presentation at the SIIM 2008 Annual Meeting in Seattle.

References

1. NEMA, Digital Imaging & Communications in Medicine (DICOM), Part 14: Grayscale Standard Display Function. 1998, National Electrical Manufacturers Association.
2. Barten, P.G.J., Contrast sensitivity of the human eye and its effects on image quality. 1999, Bellingham, Wash.: SPIE Optical Engineering Press. xix, 208.
3. Tchou, P., M. Flynn, and E. Peterson, 2AFC assessment of contrast threshold for a standardized target using a monochrome LCD monitor. Proceedings of SPIE, 2004. 5372(Medical Imaging 2004: Image Perception, Observer Performance, and Technology Assessment): p. 344-352.
4. Flynn, M.J., et al., High-fidelity electronic display of digital radiographs. Radiographics, 1999. 19(6): p. 1653-69.