Human Vision Model to Predict Observer Performance: Detection of Microcalcifications as a Function of Monitor Phosphor Elizabeth Krupinski, PhD Jeffrey Johnson, PhD Hans Roehrig, PhD Jeffrey Lubin, PhD Michael Engstrom, BS
Acknowledgments This work was supported by a grant from the NIH R01 CA We would also like to thank Siemens for the loan of 1 of the monitors and MedOptics for 1 of the CCD cameras used in the study
Rationale Digital mammography potential – Improve breast cancer detection – CAD does not need digitization Display monitors should be optimized – Physical evaluation parameters – Psychophysical evaluation (JNDs) – Clinical evaluation radiologists
Rationale Observer trials (ROC studies) – Require many images (power) – Require many observers (power) – Are time-consuming Predictive models may help – Simulate effects softcopy display parameters on image quality – Predict effects on performance
JNDmetrix Model Computational method predicting human performance in detection, discrimination & image-quality tasks Based on JND measurement principles & frequency-channel vision-modeling principles 2 input images & model returns accurate, robust estimates of visual discriminability
JNDmetrix Model
Display Monitors 2 Siemens high-performance – 2048 x 2560 resolution – Dome MD-5 10-bit video board – 71 Hz refresh rate – Monochrome – Calibrated to DICOM-14 standard P45 vs P104 phosphor
Physical Evaluation Luminance: 0.8 cd/m 2 – 500 cd/m 2 ) – Same on both NPS: P104 > P45 SNR: P45 > P104 Model input – Each stimulus on CRT imaged with CCD camera
Phosphor Granularity P45 Phosphor < P104 Phosphor
Monitor NPS
Images Mammograms USF Database 512 x 512 sub-images extracted 13 malignant & 12 benign Ca ++ Removed using median filter Add Ca ++ to 25 normals 75%, 50% & 25% contrasts by weighted superposition of signal- absent & present versions 250 total images Decimated to 256 x 256
Edited Images Original 75% Ca++ 50% Ca++ 25% Ca++ 0% Ca++
Image Editing Quality 512 x 512 & 256 x 256 versions 200 pairs of images – Original contrast only – Paired with edited version – Paired randomly with others 3 radiologists 2AFC – chose which is edited
Editing Quality Results Reader 512 x x % 46% 2 57% 47.5% 3 39% 49.5% Average 47.83% sd = % sd = 1.08
Observer Study 250 images – 256 x 5 contrasts 6 radiologists No image processing Ambient lights off No time limits 2 reading sessions ~ 1 month apart Counter-balanced presentation
Observer Study Images presented individually Is Ca ++ present or absent Rate confidence 6-point scale Multi-Reader Multi-Case Receiver Operating Characteristic* * Dorfman, Berbaum & Metz 1992
Human Results * * * * P < 0.05
Model Results * P < 0.05 * * * *
Correlation
Summary P104 – > light emission efficiency – > spatial noise due to granularity P45 – > SNR Luminance – noise tradeoff P45 > P104 detection performance JNDmetrix model predicted well
Model Additions Eye-position will be recorded as observers search images to determine if any attention parameters can be added to JNDmetrix model to improve accuracy of predictions