Automatic colorimetric calibration of human wounds

Abstract

Background: Recently, digital photography in medicine is considered an acceptable tool in many clinical domains, e.g. wound care. Although ever higher resolutions are available, reproducibility is still poor and visual comparison of images remains difficult. This is even more the case for measurements performed on such images (colour, area, etc.). This problem is often neglected and images are freely compared and exchanged without further thought.

Methods: The first experiment checked whether camera settings or lighting conditions could negatively affect the quality of colorimetric calibration. Digital images plus a calibration chart were exposed to a variety of conditions. Precision and accuracy of colours after calibration were quantitatively assessed with a probability distribution for perceptual colour differences (dE_ab). The second experiment was designed to assess the impact of the automatic calibration procedure (i.e. chart detection) on real-world measurements. 40 Different images of real wounds were acquired and a region of interest was selected in each image. 3 Rotated versions of each image were automatically calibrated and colour differences were calculated.

Results: 1st

Experiment: Colour differences between the measurements and real spectrophotometric measurements reveal median dE_ab values respectively 6.40 for the proper patches of calibrated normal images and 17.75 for uncalibrated images demonstrating an important improvement in accuracy after calibration. The reproducibility, visualized by the probability distribution of the dE_ab errors between 2 measurements of the patches of the images has a median of 3.43 dE* for all calibrated images, 23.26 dE_ab for all uncalibrated images. If we restrict ourselves to the proper patches of normal calibrated images the median is only 2.58 dE_ab! Wilcoxon sum-rank testing (p < 0.05) between uncalibrated normal images and calibrated normal images with proper squares were equal to 0 demonstrating a highly significant improvement of reproducibility. In the second experiment, the reproducibility of the chart detection during automatic calibration is presented using a probability distribution of dE_ab errors between 2 measurements of the same ROI.

Conclusion: The investigators proposed an automatic colour calibration algorithm that ensures reproducible colour content of digital images. Evidence was provided that images taken with commercially available digital cameras can be calibrated independently of any camera settings and illumination features.

Figure 1

Chronic Wound and Reference Chart. Chronic Wound and Reference Chart after (left) and before (right) calibration.

Figure 2

RGB to sRGB Transformation Scheme.

Figure 3

Chronic Wound Images with Reference Chart and a Region of Interest.

Figure 4

Example with Nikon D200: MBCCC under illuminant A (3000 K). Camera: exposure bias -1, automatic white balance.

Figure 5

Example with Nikon D200: MBCCC under illuminant A (3000 K). Camera: exposure bias -1, calibrated image.

Figure 6

Example with Canon 10D: MBCCC under illuminant A (3000 K). Camera: exposure bias -1, manual white balance.

Figure 7

Example with Canon 10D: MBCCC under illuminant A (3000 K). Camera: exposure bias -1, calibrated image.

Figure 8

Example with Nikon D200: MBCCC under illuminant D65 (6500 K). Camera: exposure bias -1, manual white balance.

Figure 9

Example with Nikon D200: MBCCC under illuminant D65 (6500 K). Camera: exposure bias -1, calibrated image.

Figure 10

Accuracy of Color Calibration. Probability distribution of dE*ab errors between the patches of the images and spectrophotometric measurements. Based on 39 images (Nikon D200) & 15 images (Canon 10D) under different illuminants and settings. Median dE*ab is 6.40 for the proper patches of calibrated normal images, 17.75 for uncalibrated images.

Figure 11

Reproducibility of Color Calibration. Probability distribution of dE*ab errors, based on 39 images taken with a Nikon D200 and 15 images with a Canon 10D under different illuminants and settings. Median of 3.43 dE*ab for all calibrated images, 23.26 dE*ab for all uncalibrated images, a median of 2.83 dE*ab for all 'normal' calibrated images and 14.25 dE*ab for all 'normal' uncalibrated images. If we restrict ourselves to the proper patches of normal calibrated images the median is only 2.58 dE*ab

Figure 12

Accuracy: boxplot for the proper patches of the normal images.

Figure 13

Reproducibility: boxplot for the proper patches of the normal images.

Figure 14

Probability distribution of dE*ab errors with region of interest calibration.