Evaluation of vermillion border descriptors and relevance vector machines discrimination model for making probabilistic predictions of solar cheilosis on digital lip photographs

Abstract

IntroductionSolar cheilosis (SC), a common precancer of the lower lip with a high potential to progress to invasive squamous cell carcinoma, presents with characteristic morphological vermillion-skin border alterations, like the border retraction. AimTo determine robust macro-morphological descriptors of the vermillion border from non-standardized digital photographs and to exploit a probabilistic model for SC recognition in real clinical environments. MethodsLip borders of 150 individuals (75 SC patients, 75 non-SC controls) were quantified on the basis of the extent of vermillion retraction and the degree of border irregularity employing fractal features and type-P Fourier descriptors. Eight lip border quantifiers were evaluated in terms of their reliability and reproducibility. The probabilistic ‘diagnostic’ model was implemented using the relevance vector machine (RVM) algorithm. ResultsPicture acquisition contributes substantially to overall variability of lip border images; however, for the different lip descriptors 33% to 65% of border morphological variability is due to differences among individuals. Different camera operators or the use of cameras with different specifications did not affect significantly the extracted lip features. The proposed RVM probabilistic model yielded a high sensitivity and specificity of 94.6% and 96%, respectively. ConclusionWe explored the use of digital photography within the clinical routine setting to validate a probabilistic model for the assessment of lip conditions like SC. The proposed method opens new perspectives toward a cost effective, non-invasive monitoring of SC to support large scale epidemiological and interventional studies in different clinical environments.