A spectroscopy-based machine learning framework on photoacoustic imaging data for automatic 3D delineation of melanoma tumors

The incidence of melanoma is rising, increasing the need for efficient diagnostics. Excisional biopsy and histopathology remain the “gold standard” but often require multiple surgical incisions due to limitations in making a visual assessment of melanoma borders either by eye or using dermatoscopy. This challenge partly arises from the inability to non-invasively produce the necessary spectral contrast to reliably delineate melanomas without chemically staining the tissue. Spectroscopic imaging provides an intriguing alternative to non-invasively characterize the molecular composition of tissue, and on that basis distinguish melanoma from healthy tissue. This work presents a computational framework for multispectral photoacoustic (PA) imaging data of melanoma skin tumors that automatically determines tumor borders without human input. The framework combines K-means clustering for training data generation, a 1D convolutional neural network to classify pixels based on their spectral features, and an active contour algorithm to delineate the tumor in both 2D and 3D. An important feature of our model is that the training data is contained within one patient and does not rely on a population-based signature, yielding an individualized diagnostic approach that is of high relevance for precision skin tumor diagnostics. Non-invasive determination of melanoma tumor borders can improve clinical practice by providing pre-surgical and perioperative guidance for complete tumor removal with minimal incisions.