Observer Performance Studies

The Influence of Image Processing on Chest Radiograph Interpretation and Decision Changes

The effect of image processing on diagnostic performance for reading computed radiography (CR) chest radiographs displayed on a workstation monitor was investigated. Of special concern was an analysis of decision changes occurring from viewing images without processing to viewing them with processing. The two major questions were 1) whether image processing actually leads to an increase in the true-positive rate, and 2) was there any correlation between the type of image processing used and the type of lesion present on the image [5].

Methods

Images

  • 168 CR Chest Images
  • 38 with subtle instances of nodule
  • 51 with interstitial disease
  • 15 with bronchiectasis
  • 64 lesion free

Image Processing Functions

  • Linear (similar to film)
  • Lung Window
  • Mediastinal Window
  • Black & White Reverse
  • Low-Pass Filter
  • High-Pass Filter

Display

  • 2k x 2.5k monitor
  • Perceptually Linearized

Procedure
6 Radiologists Viewed each image first without any processing (using the linear setting) Reported present or absent for each of the 3 diagnostic categories Reported confidence using 6-level scale Used any of the image processing functions avaliable Revised initial decisions & confidences At the end of the study each observer subjectively rated each processing function as excellent, good, fair or poor

Results

ROC Performance Analysis

  • Mean ROC Az without presets = 0.8558
  • Mean ROC Az with presets = 0.8536
  • 95% confidence limits on the difference (-0.0195, 0.0239)
  • df = 10, F = 0.0507, p = 0.8262
  • No significant difference in performance with & without image processing

Decision Changes

  • Percentages of types of decisions made without image processing and with image processing. The first 4 categories represent those in which no decision change occurred. The last four categories represent those where decisions did change. TP = true-positive; FP = false-positive; FN = false-negative; TN = true-negative.
Without Processing With Processing % of Cases
TP TP 37
TN TN 28
FP FP 7
FN FN 21
TP FN 2
FN TP 2
FP TN 1
TN FP 2

Average Number of Times per Image Processing Used

  Normal Nodule Interstitial Bronchiectasis
Linear 1.27 1.28 1.34 1.27
Lung 0.95 0.93 1.03 1.10
High-Pass 0.56 0.54 0.52 0.56
Reverse 0.49 0.53 0.50 0.56
Mediastinum 0.41 0.44 0.41 0.41
Low-Pass 0.05 0.05 0.06 0.01

Conclusions

  • There were no statistically significant differences in detection performance without & with use of image processing
  • Negative and positive decision changes without & with image processing balanced each other out - resulting in no gain in
    performance
  • There was no correlation between the type of image processing used and the type of lesion on the image
  • The linear, lung and high-pass processing functions were used most frequently

Return to Observer Performance Studies page