Вы находитесь на странице: 1из 1

Program # 3353

Investigating Mean Pixel Height Standard Deviation: An Image Quality Board # B886
Metric Used in Scanning Laser Tomography
N O’Leary,1 AJ Patterson,1 DF Garway-Heath, 2 DP Crabb1
1Department of Optometry and Visual Science, City University London, London, UK
2Glaucoma Research Unit, Moorfields Eye Hospital, London, UK

INTRODUCTION 1 2 3 73 74 RESULTS
z-axis Original Single

Mean Pixel Height Standard Deviation (MPHSD) is a metric to Topography Series • The range of MPHSD Values for generated MTs (averaged over 10 simulations)
measure the quality of alignment of single topographies composing a was very large (15 to 114μm) for identical misalignment noise sets
Identical
Heidelberg Retinal Tomography (HRT) Mean Topography (MT) image. Generated Single
Topographies
Noise Sets
• PHSD values for generated MTs were not Normally distributed (Figures 3 and 4)
This measure is frequently the main factor in deciding if a MT image is and were positively skewed with larger tails in ONH structures with higher
Generated Mean
to be included in studies and/or if further scans should be taken to y-axis Topographies
x-axis topographic height differences between the bottom of the cup and the rim
improve the image quality.
• As a consequence, MPHSD is sensitive to misalignment in images with more
This study used HRT patient data and computer simulation to assess:
Figure 1. Schematic of single series simulation (as applied to 74 distinct single prominent anatomical features, primarily a higher degree of cupping
• The sensitivity of MPHSD of HRT MT images to the qualitative
topographies) with noise set visualisation (randomly generated translations in, • The median of PHSD is less sensitive to these effects of ONH structure (Figure 4)
structure of the constituent single topographies
and rotations about, the x, y, z axes, with pixel-wise Gaussian noise added2)
• The sensitivity of other metrics, such as the Median of PHSD of HRT • Qualitative observations of images reveals also that the MPHSD is most sensitive
MT images to high frequency spatial variability in topographic images (Figure 5)
Heidelberg Engineering have published guidelines for interpreting
MPHSD
try to Cross section of Topographies
RATING perfect very good good acceptable not evaluable
improve
giving 3 minimum MPHSD values
MPHS (μm) < 10 10-20 20-30 30-40 40-50 > 50
Figure 2. HRT topography images (with reflectance images) giving minimum
and these have been used for inclusion/exclusion criteria in several
(left) and maximum (right) MPHSD values (image structures to which MPHSD is
studies e.g. 3. least and most sensitive, respectively, of the 74 images when uniform noise is Cross section of Topographies
added) giving 3 maximum MPHSD values
METHODS
Generated Mean Topography:
Simulation generates 3 single topographies from one original single Higher PHSD Figure 5. Cross-section of topographies representing low-frequency spatial variability
and high-frequency spatial variability.
topography, then averaged at a pixel level to produce a ‘new’ MT.
Calculating MPHSD:
At each pixel in a MT a Pixel Height Standard Deviation (PHSD) is CONCLUSIONS
calculated over the 3 constituent Single Topography pixels. The • The range of MPHSD values varied considerably despite identical misalignment
arithmetic mean over all PHSD in the MT is the MPHSD. movement being applied to generate all mean topographies
Simulation Methodology • As the distribution of pixel height standard deviation values is skewed, the mean
74 Single Topographies from may not be an appropriate summary measure to characterise the data
3 Noise Sets
real patient data1 Figure 3. 2-D and 1-D distributions of PHSD for generated MTs of those • The median of PHSD is less sensitive to misalignment of topographies with high
giving minimum (left) and maximum (right) MPHSD values
anatomical structural variability
74 Generated Mean 74 Associated MPHSD • Both MPHSD and median of PHSD are sensitive to misalignment of topographies
Topography
Topographies Values Topography
with local high frequency variability
• Developed by Patterson et al2 to provide realistic virtual patients to
aid in developing and testing measurement methods FURTHER WORK
•Carefully engineered to produce MPHSD values consistent with the • We plan to evaluate other metrics to summarise the repeatability of topography
distribution of those obtained clinically images and compare them with image quality measures derived from expert
• Uses real patient data (single topographies) and simulated observers
misalignment noise to generate MTs
• For a series of 74 single topographies identical misalignment noise 1. Strouthidis et al: Br. J. Ophthalmol. 2005; 89; 1427-1437 2. Patterson et al: IOVS; 2005; 46; 1657-1667
3. Hawker et al: IOVS; 2005; 46; 4153-4158
was used to produce 74 MTs (Figure 1) Figure 4. 2-D and 1-D distributions of PHSD for generated MTs of those
• In total 10 MTs, each with different noise, were generated for each of giving MPHSD values of 36 (left) and 27μm (right) but Median of PHSD values This work is supported in part by an unrestricted educational grant from
the 74 single topographies along with the associated MPHSD values. both of 22μm. Note the longer tail in the left distribution to which the Heidelberg Engineering and the Moorfields NHS Special Trustees
• Individual PHSD values were examined in order to examine the pixel- arithmetic mean is more sensitive than the median.
Author disclosure block: N O’Leary: None; AJ Patterson: None; DF Garway-Heath: Carl Zeiss Meditec
wise distribution of standard deviations. Inc.: F, Talia Technologies Ltd: F, Heidelberg Engineering: F, R ; DP Crabb: None.

Вам также может понравиться