FPGA High Level Synthesis for the Classification of Skin Tumors with Hyperspectral Images

Cancer is the main cause of premature death in the world, with 18 million diagnoses in 2018, 3.9 million of which in Europe. In particular, according to studies conducted by the American Academy of Dermatology, skin cancer is the most prevalent type in the US. Diagnostic tools are generally invasive, hence research focuses on emerging technologies, like hyperspectral images, since they are non-invasive, contactless and non-ionizing. A hyperspectral image acquisition system has been used to produce a database of 49 images from 36 patients, used to validate an innovative machine learning algorithm. Starting from our original serial implementation, a novel version has been developed with modern technologies of High Level Synthesis (HLS) using FPGA, to verify the feasibility of a portable instrument. Differences in implementation, HLS optimizations and latency times have been compared and evaluated. The algorithm has been tested on different FPGAs, to identify the optimal device for the purpose. Finally, the proposed hardware architecture processes hyperspectral images dissipating less energy than state-of-the-art GPU solutions.