Advancing non-invasive melanoma diagnostics with deep learning and multispectral photoacoustic imaging.

The incidence of melanoma is rising and will require more efficient diagnostic procedures to meet a growing demand. Excisional biopsy and histopathology is still the standard, which often requires multiple surgical incisions with increasing margins due inaccurate visual assessment of where the melanoma borders to healthy tissue. This challenge stems, in part, from the inability to reliably delineate the melanoma without visually inspecting chemically stained histopathological cross-sections. Spectroscopic imaging have shown promise to non-invasively characterize the molecular composition of tissue and thereby distinguish melanoma from healthy tissue based on spectral features. In this work we describe a computational framework applied to multispectral photoacoustic (PA) imaging data of melanoma in humans and demonstrate how the borders of the tumor can be automatically determined without human input. The framework combines K-means clustering, for an unbiased selection of training data, a one-dimensional convolutional neural network applied to PA spectra for classifying pixels as either healthy or diseased, and an active contour algorithm to finally delineate the melanoma in 3D. The work stands to impact clinical practice as it can provide both pre-surgical and perioperative guidance to ensure complete tumor removal with minimal surgical incisions.