Deep learning model deploying on embedded skin cancer diagnostic device

The number of research papers, where neural networks are applied in medical image analysis is growing. There is a proof that Convolutional Neural Networks (CNN) are able to differentiate skin cancer from nevi with greater accuracy than experienced specialists on average (sensitivity 82% and 73% accordingly).1 Team's latest research2 allows achieving even greater accuracy, by using specific narrow-band illumination. Nevertheless, the overall probability of early skin cancer detection depends on the availability of diagnostic tools. If screening tools will be available to a high number of general practices, the chance of disease detection will increase. The previous research3 shows that scalable cloud service is able to process a high number of users. After a certain number of users, the overall cost of the system, including cloud processing expenses and cost of high computational power portable device, might be higher if compared to an on-premises solution, where each device is capable of diagnosing without Internet access. It might be cheaper to equip devices with additional neural processing unit (NPU) and exclude cloud processing. Another option is to make screening available by using the newest smartphones that are equipped with NPU.4 The problem of using the NPU is that they are limited in storage space, accuracy, and features. Therefore, a full-size CNN model should be adapted and minimized to fit in a limited NPU. Research reviews existing CNN optimization methods and proposes the most accurate for skin cancer diagnostics. The paper evaluates CNN prediction losses when the model's elements’ precision is reduced from 32 bits to 8 and rounded to integer values.