A Machine Learning Approach to Skin Cancer Delineation on Photoacoustic Imaging

Abstract: Skin cancer is a growing public health concern due to its prevalence among the population. Current clinical procedures require high invasiveness and multiple surgeries, which are responsible for patient discomfort and high medical expenses. Photoacoustic imaging offers an alternative to standard skin carcinoma diagnosis by exploiting the low scattering rate of ultrasound, which enables deep tissue penetration, and high imaging resolution. This thesis focuses on the application of machine learning models to systematically identify and delineate tumour regions according to their photoacoustic spectra. This is achieved through the implementation of a multilayer perceptron and a convolutional neural network that binary classify the spectral inputs. The same procedure is repeated on data that were dimensionally reduced by an autoencoder. The model predictions are further refined through post-processing contouring techniques. We apply our approach to six samples of the most common skin tumour types and successfully estimate the carcinoma extension. Specifically, the convolutional network accurately estimated the tumour extension, this way consistently decreasing the size of excision margins. This model combined with the contouring technique constitutes a safe approach to skin cancer diagnosis. Our method significantly reduces the number of required surgeries and once automated will decrease the current medical staff workload.