Machine learning for optical skin microscopy: a tutorial, current advances, and challenges (Conference Presentation)

Abstract

Traditional clinical diagnosis and research on skin lesions, performed by visual examination, dermoscopy, biopsy, and pathology, are complemented with newer noninvasive optical imaging approaches, including reflectance confocal microscopy (RCM) and optical coherence tomography (OCT). Limitations such as single contrast (gray-scale) (RCM, OCT), limited structure-specific contrast (RCM, OCT), en face orientation (RCM), relatively low resolution (OCT), and spatially variable speckle noise (RCM, OCT), contrary to the orthogonal orientation, purple and pink color contrast and noise-free appearance of pathology. Interpreting nuclear, cellular and morphologic patterns at different magnifications and scales are mostly manual, qualitative and subjective, with consequent intra- and inter-observer variability among experts and extensive training requirements for novices. These new / developing approaches need quantitative, accurate and repeatable image reading and analysis tools, which may be created with machine learning (ML) and associated methods. Recent advances in ML offer an intellectually rich sandbox, which can be simply and naively applied as off the shelf solutions. However, we contend that for longer-term success, it is critical to avoid such off-the-shelf solutions and instead design novel, specialized, microscopy-specific ML algorithms. The sandbox provides ideas, concepts, developments, low-level feature extraction tools and higher-level ML tools. Recent work has focused on using ML for detection of the dermal-epidermal junction in image-stacks (RCM, OCT), classification of cellular patterns in image-mosaics of melanocytic lesions (RCM), basal cell carcinoma detection (OCT), and videomosaicking (RCM). In this presentation, we provide a tutorial on applying ML to skin microscopy in the context of our experience developing novel learning models for RCM skin image analysis.