Effects of Grayscale WindowLevel on Breast Lesion Detectability

1
Effects of Grayscale Window/Level on Breast
Lesion Detectability

• Jeffrey Johnson, PhD a
• John Nafziger, PhD a
• Elizabeth Krupinski, PhD b
• Hans Roehrig, PhD b

b
a
Supported by U. S. Army Medical Research and
Materiel Command, grant DAMD-17-01-1-0621
2
Rationale

• Nearly 50 of breast lesions missed at initial
  screening are visible retrospectively
• Digital mammography could reduce perceptual
  errors by enhancing lesion conspicuity with image
  processing
• Perceptual models could be useful tools for
  automating and optimizing techniques for image
  enhancement

3
Overview

• This study evaluated the use of a visual
  discrimination model (VDM) for predicting effects
  of one type of image enhancement - grayscale
  window width and level (W/L) - on the
  detectability of breast lesions
• Compared model and observer performance in two
  experiments
• 2AFC detection thresholds with simulated
  mammograms and nonmedical observers
• ROC observer performance study with radiologists
  and digitized mammograms

4
Methods Simulated Mammograms

• Backgrounds
• Filtered noise, 1/f3 noise power spectrum
• Two groups Bright and Dark central regions

• Lesion signals
• Mass 2D Gaussian (d50 arcmin)
• Microcalcification cluster six blurred disks or
  specks (disk d8 arcmin, cluster d40 arcmin)

5
Methods W/L Conditions

• P-value transformations
• Fully stretched
• Understretched (-25)
• Overstretched (25)
• Bright shifted (25)
• Dark shifted (-25)
• Applied to full 512x512 pixel image or 170x170
  pixel central region of interest containing lesion

6
Example Test Images
Specks Full W/L Dark Center
Specks Central W/L Dark Center
Gaussian Full W/L Bright Center
Gaussian Central W/L Bright Center
Fully stretched (FS)
Under stretched (US)
Over stretched (OS)
7
Example Test Images
Gaussian Full W/L Bright Center
Gaussian Central W/L Bright Center
Specks Full W/L Dark Center
Specks Central W/L Dark Center
Bright shifted (BS)
Dark shifted (DS)
8
2AFC Threshold Detection

• Side-by-side presentation of same background
  with/without signal
• Signal amplitude varied in 1-up/3-down staircase
  procedure detection threshold at 80 correct
• Five W/L conditions interleaved in same session
• Separate sessions for two signal and two
  background types

9
Test Conditions

• Siemens 5M-pixel CRT monitor (P45)
• Luminance range 0.3 to 290 cd/m2
• Barco 10-bit display controller
• DICOM-14 grayscale display function
• Three nonmedical observers
• Viewing distance 52 cm chin rest
• Ambient lights off

10
Results Detection Thresholds for Gaussian Signals
Bright Backgrounds
Dark Backgrounds
Error bars show 95 confidence intervals
11
Results Detection Thresholds for Speck Clusters
Bright Backgrounds
Dark Backgrounds
Error bars show 95 confidence intervals
12
Experimental Detection Thresholds

• Significant variations across W/L conditions
• Generally lower for central vs. full W/L
• due to local contrast enhancement
• - fully stretched not always optimal
• Full W/L Lowest thresholds for
• fully stretched, understretched (specks only)
• dark shifted on bright, bright shifted on dark
• Central W/L Lowest thresholds for
• overstretched for Gaussians and specks on dark
• dark shifted on bright, bright shifted on dark

13
Visual Discrimination Modeling

• Simulates physiological response of human visual
  system to visual stimuli luminance patterns from
  images video
• Output is a deterministic prediction of feature
  or image discriminability as function of spatial
  location, spatial frequency, and time
• Discriminability measured in units of Just
  Noticeable Differences (JND)

14
VDM Architecture
Contrast Pyramid (visual cortex)
Pair of input images
Spatial orientation responses
Display luminance
Within-band Masking

Spatial frequency bands
Crossband Masking

Optics
Contrast Pyramid
JND Distance
Probability
Display Ocular Processing
Combin. Rule
JND scalar
JND map
15
VDM vs. Experimental Thresholds for Gaussians on
Bright Backgrounds
Full W/L
Central W/L
Error bars show 95 confidence intervals
16
VDM vs. Experimental Thresholds for Gaussians on
Dark Backgrounds
Full W/L
Central W/L
Error bars show 95 confidence intervals
17
VDM vs. Experimental Thresholds for Specks on
Bright Backgrounds
Full W/L
Central W/L
Error bars show 95 confidence intervals
18
VDM vs. Experimental Thresholds for Specks on
Dark Backgrounds
Full W/L
Central W/L
Error bars show 95 confidence intervals
19
VDM vs. Experimental Thresholds Simulated
Lesions Backgrounds

• Generally good agreement between model and
  experimental detection thresholds and variations
  across W/L conditions
• Consistently reduced thresholds with central
  (local ROI) vs. full-image W/L
• Largest modeling discrepancies for specks,
  especially on dark backgrounds

20
ROC Observer Study

• Determine effects of W/L functions and size on
  detection of microcalcification clusters by
  mammographers
• Evaluate utility of localized ROI contrast
  enhancement (central vs. full W/L)

21
ROC Observer Study Image Preparation

• Digitized mammograms (n15) from Digital Database
  for Screening Mammography
• Extracted 512x512-pixel sections with single,
  centered microcalcification cluster
• Removed calcifications by median filtering
• Generated five lesion-contrast levels 0, 25, 50,
  75, and 100
• Applied three W/L functions Fully stretched,
  under and over stretched by 15
• Full and Central W/L sizes

22
ROC Observer Study Test Conditions

• 6 radiologists at Univ. of Arizona
• 225 images/session
• 2 reading sessions 2 weeks apart
• Decision confidence on 6-point scale
• No image processing, no time limits, ambient
  lights off viewed at 25 cm
• Siemens 5M-pixel CRT monitor (P45)
• Luminance 0.8 to 500 cd/m2
• DICOM-14 grayscale display function

23
Examples of Test Images
Overstretched (OS, 15)

• Fully stretched
• (FS, 0-4095)

Understretched (US, 15)
Full W/L
Central W/L
24
ROC Observer Study Results

• Compared central vs. full W/L across all W/L
  functions, all lesion contrasts
• Observer performance statistically better
  (plt0.05) for FULL W/L size

Az Values
25
ROC Observer Study Results

• No statistically significant variations
• between central and full W/L sizes for a single
  W/L function (all lesion contrasts)
• between central and full W/L sizes for a single
  combination of W/L function and lesion contrast
  (except FS, 50)
• across W/L functions in central and full W/L
  sizes considered separately (all lesion contrasts)

26
ROC Observer Study Analysis

• Central W/L enhanced lesion contrast but changed
  appearance of parenchymal tissue relative to
  surrounding areas
• Decision confidence lowered by nonuniform
  appearance of background tissue characteristics
• Conclusion Calcifications may be easier to
  perceive (due to higher contrast) but more
  difficult to interpret (due to cognitive factors,
  past experience)

27
Conclusions

• For simulated lesions and backgrounds, VDM was
  generally a reliable predictor of W/L conditions
  for optimal detectability
• Results with simulated images suggested benefits
  of localized contrast enhancement
• Decision confidence and performance of
  mammographers actually lower with localized W/L,
  probably due to nonuniform tissue appearance

28
Future Directions

• Allow toggling between full and local W/L modes
  (combine uniform contextual data with local
  contrast enhancement)
• Evaluate effects of W/L on detection of very
  subtle lesions (low contrast, near threshold)
• Model refinements
• improved crossband masking for higher frequency
  signals specks/calcifications
• include effects of background noise via
  statistical observer model