A comparative study revealed that color constancy algorithms reduce the variability of dermoscopy image quality, leading to positive outcomes for practitioners.
Artificial intelligence (AI)- based color constancy algorithms, when applied to dermoscopical assessment of skin lesions, effectively improved the quality of dermoscopic imaging and led to positive outcomes for practitioners, according to a comparative study published in Skin Research and Technology.1
Investigators Branciforti et al sought to explore the efficacy and impacts of an AI-based color constancy algorithm known as DermoCC-GAN in the context of skin checks and lesion diagnosis. They noted that while it has been established that these algorithms improve the quality of dermoscopic images by reducing variability in factors such as lighting, there is limited research exploring the efficacy of the algorithm's impact on practitioner workflow.
Specifically, the study explored the algorithm's impacts on image quality as perceived by dermatologists, the patient's diagnosis, and the dermatologist's confidence in the diagnosis.
Using an open access data set, investigators started with a total of 150 dermoscopic images of 5 different lesion types, including actinic keratosis, basal cell carcinoma, keratosis-like lesions, melanoma, and nevus. Lesion types were included in the study based on clinical setting frequency and potential for diagnostic challenges due to a diversity of morphological characteristics.
All 150 images were processed with DermoCC-GAN, leading to a total of 300 images included in the study in total, with 150 original images and 150 images processed by the algorithm.
Three dermatologists with various levels of clinical experience participated in the study, carrying out both an unpaired evaluation task and a paired evaluation task.
In the unpaired evaluation task, the dermatologists evaluated the images for overall image quality, lesion diagnosis, and diagnosis confidence. In the paired evaluation task, the dermatologists evaluated the images for impact of normalization, lesion diagnosis, and diagnosis confidence.
Regarding the image quality evaluation, dermatologists examined brightness, sharpness, and chromatic components. In total, images normalized by the AI-based color constancy algorithm were visually perceived to be of higher quality than the original images.
"It is evident that the use of normalized images by AI-based color constancy algorithms, such as DermoCC-GAN, brings qualitative benefits to the clinical practitioner on skin lesions diagnostic routine. However, acknowledging the importance of extreme caution in this area, we always suggest the simultaneous analysis of original and normalized images," wrote study authors. "This approach enables dermatologists to extract essential information from both images, contributing to a more accurate classification. The combination of the original and normalized images not only enhances the diagnostic capability of the clinicians but also strengthens their confidence level during the diagnostic process."