Automated scoring of vitiligo using superpixel-generated computerized digital image analysis of clinical photographs: a novel and consistent way to score vitiligo

Dear Editor, Objective outcome measures for vitiligo are essential for evaluation of severity and response to treatment. The Vitiligo Area Scoring Index (VASI) is a validated tool to assess the severity of vitiligo.1,2 However, estimation of the area and extent of depigmentation in VASI is often subject to much inter‐ and intraobserver variability.3 We aimed to develop and validate novel image analysis software designed to calculate, automatically, areas of depigmentation in vitiligo from computer image analysis of clinical photographs, thereby deriving automated VASI scores. The proposed algorithm was developed in collaboration with image analysis experts. It derives the score in three steps. Firstly, the background and undergarments of each image are identified and segmented away using segmentation methods. Next, the images are converted into superpixels using linear spectral clustering.4 These superpixels are then used for feature extraction and further classification of vitiligo vs. nonvitiligo areas. Finally, the algorithm expresses the segmented vitiligo area as a percentage of the total body surface area (BSA). Each image is processed in 40 s. Percentage BSA involvement is subsequently validated in a prospective clinical trial. During validation, 21 patients with vitiligo undergoing narrowband ultraviolet B therapy were recruited. VASI scores were obtained during clinical visits at baseline and 3 and 7 months after treatment initiation. A set of 10 standardized photographs focusing on the trunk and upper and lower limbs was taken at the same time point. In this study, only trunk photos (front and back, left and right lateral) were used and sent for digital image analysis to determine percentage BSA involvement of the trunk. Algorithm‐derived VASI (aVASI) values were compared with both VASI(c) score – done during clinical visits – and VASI(p) score – derived from clinical photographs by dermatologists. The results showed that our computerized algorithm produced estimates of skin depigmentation that were correlated significantly with both VASI(c) and VASI(p), with correlation coefficients (r) > 0·7 and P‐values < 0·001. aVASI correlated better with VASI(p) (r = 0·97) than with VASI(c) (r = 0·72). There was a good, statistically significant correlation when the scores were analysed at the individual time points of baseline, after 3 months and after 7 months (Table 1). Correlation of aVASI, VASI(p) and VASI(c) at and between baseline, 3 months and 7 months VASI, Vitiligo Area Scoring Index. aVASI, algorithm‐derived VASI; VASI(p), VASI scored from clinical photographs; VASI(c), VASI scored from clinical visits. Correlation of aVASI, VASI(p) and VASI(c) at and between baseline, 3 months and 7 months VASI, Vitiligo Area Scoring Index. aVASI, algorithm‐derived VASI; VASI(p), VASI scored from clinical photographs; VASI(c), VASI scored from clinical visits. Changes in scores of aVASI and VASI(p) over different time points correlated strongly (r = 0·7–0·9, P‐values < 0·001). In comparison, changes of aVASI and VASI(c) between these time points appeared less consistent (r = 0·3–0·7). VASI(p) and VASI(c) appear to correlate only moderately, showing inconsistencies in manually scored VASI. While these two VASI scores correlated well at each time point (r = 0·5–0·9), changes of these scores between time points appear less consistent (r = 0·2–0·7). Scoring of VASI during actual clinical visits may be less accurate, as doctors running trials may have limited time to score the extent and degree of depigmentation accurately. Correlations between aVASI scores and VASI(c) and VASI(p) appear less consistent when the extent of skin depigmentation increases. Compared with smaller areas of vitiligo, manual estimation of the extent of depigmentation by the palmar method is harder, and variability in the degree of residual depigmentation is greater in larger areas. These factors result in more human error. The algorithm can reflect small changes in areas of involvement, as the extent of involvement is expressed as a continuous variable. This is different from conventional systems, which score the degree of depigmentation in bands. The ability to record small changes is important when assessing stability of disease and response to treatment in the interim of clinical trials, especially when comparing different treatment options and regimens. The algorithm is novel as it combines the use of global and adaptive threshold segmentation techniques and superpixel methods to select and calculate vitiligo areas accurately. Using linear spectral clustering to convert input images into superpixels can help identify and classify pixels near boundaries of normal and vitiliginous skin to obtain higher degrees of image segmentation. This improves the accuracy in calculating depigmentation. This technique has not been used in medical imaging. This algorithm proposes an approach in vitiligo scoring using digital image analysis by producing automated scores, which show statistically significant strong correlations with conventional manually scored VASI. It reduces intra‐ and interassessor variability by leveraging on digital analysis of clinical photographs. As scoring is done automatically, no prior training is needed for the assessor. Our algorithm holds great promise as a fast, consistent and reliable tool that would enable clinicians around the globe to standardize and reproduce vitiligo severity scoring easily, and to compare different outcome measures and treatment options accurately.