Annu Int Conf IEEE Eng Med Biol Soc. 2022 Jul;2022:2639-2642. doi: 10.1109/EMBC48229.2022.9871108.
This work explores the possibility of applying edge machine learning technology in the context of portable medical image diagnostic systems. This was done by evaluating the performance of two machine learning (ML) algorithms, that are widely used on medical images, embedding them into a resource-constraint Nordic nrf52840 microcontroller. The first model was based on transfer learning of the MobileNetVI architecture. The second was based on a convolutional neural network (CNN) with three layers. The Edge Impulse platform was used for training and deploying the embedded machine learning algorithms. The models were deployed as a C++ library for both, a 32-bit floating point representation and an 8-bit fixed integer representation. The inference on the microcontroller was evaluated under four different cases each, using the Edge Impulse EON compiler in one case, and the Tensor Flow Lite (TFLite) interpreter in the second. Results reported include the memory footprint (RAM, and Flash), classification accuracy, time for inference, and power consumption.
本工作探讨了在便携式医疗图像诊断系统的背景下应用边缘机器学习技术的可能性。这是通过评估两种机器学习 (ML) 算法的性能来实现的，这两种算法广泛应用于医学图像中，并将其嵌入到资源受限的 Nordic nrf52840 微控制器中。第一个模型基于 MobileNetVI 架构的迁移学习。第二个模型基于具有三个卷积层的卷积神经网络 (CNN)。Edge Impulse 平台用于训练和部署嵌入式机器学习算法。这些模型被部署为 C++库，用于 32 位浮点数表示和 8 位固定整数表示。在四种不同情况下评估了微控制器上的推理，其中一种情况下使用了 Edge Impulse EON 编译器，另一种情况下使用了 TensorFlow Lite (TFLite) 解释器。报告的结果包括内存占用 (RAM 和 Flash)、分类准确性、推理时间和功耗。
