Tiny machine learning for hyperspectral skin cancer image classification

As the complexity of Machine Learning (ML) models grows, the need for powerful and high-memory devices is becoming more and more pressing. Furthermore, some scenarios like medical environments frequently deal with high-dimensional data such as clinical images that are challenging to analyze and elaborate using ML models without suitable hardware. This is the case, for instance, of hyperspectral (HS) images offering a wealth of spectral information that can aid in early detection and characterization. Many studies present in the literature address the use of Graphical Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs) for elaboration and classification of HS images in both the medical and non-medical field obtaining relevant results at the expense of a high-power consumption and a long time for designing. This paper aims at introducing an emerging subfield of ML called Tiny Machine Learning (Tiny ML) including hardware and software able to perform data analysis on extremely low-power devices such as microcontrollers, typically in the mW range or below. Furthermore, the research described in this paper wants to highlight how complex models like Artificial Neural Networks (ANNs) can be executed on tiny devices such as the Arduino Nano 33 BLE (Bluetooth Low Energy) Sense module. Results show the effectiveness of this new branch of Artificial Intelligence (AI) in classifying HS skin cancer images.