Machine Learning and Reflectance Confocal Microscopy for Biopsy-free Virtual Histology of Squamous Skin Neoplasms

Despite improvements in non-invasive medical imaging to aid in the diagnosis of internal malignancy, improvements in imaging the skin non-invasively have been slower. The dermatoscope, a device that gives a magnified and polarized view of the skin, is the only ancillary tool commonly used for clinical assessment by dermatologists to assist in diagnosis. Keratinocyte carcinomas, basal cell carcinoma and squamous cell carcinoma, are by far the most common cancers diagnosed in the United States. Due to sun exposure during military deployment, our nation’s Veterans have an increased likelihood of developing these and other skin cancers compared to the general population. Early keratinocyte carcinomas are often difficult to distinguish clinically from irritated/inflamed precancerous or benign skin lesions (actinic or seborrheic keratoses). A non- invasive technology that can assist dermatologists obtain a diagnosis of skin lesions may prevent unnecessary biopsy, resulting in fewer scars, as well as allow diagnosis and definitive treatment of skin malignancies in the same clinic visit, improving clinical workflow and patient access to dermatology clinics. A recently approved skin imaging technology, reflectance confocal microscopy (RCM), provides state-of-the-art cellular level resolution of the skin without biopsy, but still has many limitations, limiting its utility to only the most skilled users. We recently began using software-based digital enhancements to autofluorescence of unstained frozen tissue sections of microscopic slides to virtually stain unfixed tissue and provide rapid histology quality images without requiring the laborious tissue processing required of actual processing. Our overarching hypothesis is that we can apply our digital technology to overcome many of the technical limitations of RCM, and improve the dermatologists’ or pathologist’s ability to obtain more accurate diagnosis of skin lesion by RCM without requiring skin biopsy. Our preliminary data demonstrates that our software algorithms can digitally enhance RCM images of normal skin and basal cell carcinoma, resulting in histologic quality images. In Aim 1, we will use methodological and computational approaches to refine tissue processing and data acquisition to provide optimal registration of skin images to obtain the highest quality data sets to train the machine learning algorithm. In Aim 2, we will incorporate inflamed and uninflamed seborrheic keratosis, actinic keratoses, and squamous cell carcinoma skin lesions to incorporate features of these lesions into our algorithms originally developed for normal skin and basal cell carcinoma. In Aim 3, we will perform a pilot study to test the optimized virtual histology algorithm by prospectively collecting images of consecutive skin lesions on a variety of patient samples. We will compare how novice and expert RCM dermatology and pathology users perform in obtaining diagnosis using RCM with and without the virtual histology algorithm. If successful, these studies will provide an initial step towards noninvasive diagnosis of skin cancer for Veterans and civilians.