Introduction: Advancements in science and technology have propelled the surging popularity of wearable devices across various domains including human-computer interaction, motion capture, and biomedical applications. Amidst the rapid expansion of the wearable device market, the augmentation of external devices' functionality and the pursuit of elevated living standards have steered the evolution of wearable electronics towards multifunctional systems. Presently, the predominant wearable devices manifest in the form of wristbands, watches, or glasses, predominantly comprised of rigid components, with flexible elements in direct contact with the human body. In order to enhance user comfort, elevate human engagement, and achieve sophisticated healthcare outcomes, wearable electronic devices are progressing towards platforms that exhibit complete flexibility, stretchability, and even possess self-healing capabilities. Substantial strides have been made in the realm of functional flexible materials, laying the groundwork for flexible and wearable electronics. These materials have undergone significant advancements over the past decade. Drawing upon these breakthroughs, a versatile wearable imaging system powered by cutting-edge machine learning technology has been developed to detect breast cancer cells. [1] This research endeavor aims to leverage state-of-the-art wearable imaging technology and machine learning-driven data services to facilitate sensor imaging and optimize pre- and post-operative care decisions for breast cancer in the foreseeable future.
Materials and
Methods: As illustrated in Fig1a, the wearable sensing system is an intricately integrated device formed by magnetoelectric nanoparticles (MENP) based sensors and a PCB. This system is designed to detect and differentiate signals emanating from cancerous and normal cells. During the process of body scanning, alterations in the spin-echo signal were noted in correspondence to the dimensions of the cancer cells. Leveraging state-of-the-art MENP sensors, the wearable imaging system enables real-time image acquisition, with the controller employing ATSAMV71Q21B from JS Nanotechnologies for sensor signal processing. [2] The acquired sensor signals undergo ADC processing and are subsequently transformed into DAC output amplitudes to facilitate instantaneous updates. Additionally, machine learning algorithms are employed to ensure the precision of the data. Throughout the measurement procedure, a breast phantom was employed as the sample tissue, with healthy and cancerous cells being measured separately (control and experimental groups, respectively). The pMAG sensors, processing microchip, and DAC outputs collectively prepare and process the data for machine learning tasks, converting it into tensors. During the data collection phase, signals influenced by measurement inaccuracies or external variables are filtered out due to their unreliability. As the convolutional neural network undergoes training across numerous epochs, the program's accuracy steadily enhances. Through initial manual training, the system gradually becomes more proficient in autonomously organizing data over successive iterations, ultimately refining itself without the need for further manual intervention.
Results, Conclusions, and Discussions: A multitude of signals were collected to analyze the distinctions between cancerous and healthy cells. As depicted in Fig 1b, there were seemingly normal signals without errors, alongside signals exhibiting numerous measurement errors or extraneous variables. Fig 1b also highlights various inaccurate measurements, which can be rectified by the machine learning algorithm. The CNN (Convolutional Neural Network) is specifically designed to recognize and categorize both accurate and erroneous data and effectively reduce the data dimension while retaining its essential features. [3] Our data comprises a 2D tensor. To extract the primary features from the plotted data, a CNN is applied since it can be a crucial feature from the plotted images. Our CNN architecture includes three convolutional layers with one pooling layer, in addition to three fully connected layers with ReLU activation as illustrated in Fig 1c. [4] Considering we have a 2D tensor, which is 1 x 952 x 2 (channel x height x width), the initial input channel for the first layer is 1, with the feature maps expanding to 4 channels and subsequently to 32 channels before converging to 2 channels. The data is then flattened and fed into the subsequent fully connected stage. In the fully connected layers, the first layer comprises 128 neurons, which then connect to the subsequent 32 neurons layer, followed by a non-linear combination resulting in a single output of 0 or 1. As shown in Fig 1d, the loss converges to a sufficiently small value of 1.4139804e-18. Table 1 showcases the hyperparameters utilized during the training phase. Once an inaccurate measurement is identified, the data goes through another stage for correction of inaccuracies. It then automatically aligns the reference data with the signal data based on the learning from prior training. This process holds promise for optimizing data measurement in the future, thereby ensuring the device attains a high level of accuracy and efficacy. The enhancement of the machine learning algorithm utilized has the potential to rectify data measurement inaccuracies and achieve remarkably precise data identification. Further enhancements are geared towards seamlessly incorporating machine learning algorithms into wearable devices.
Acknowledgements (Optional): The authors acknowledge JS Nanotechnologies for providing the hardware to generate the data and also thank to Dr. Hang Chang (Bioscience division) at Lawrence Berkeley National Lab for his supervision.
References (Optional): [1] Song, X., Yi, B., Chen, Q et al., Machine Learning-Powered Ultrahigh Controllable and Wearable Magnetoelectric Piezotronic Touching Device. ACS Nano 2024 18(26), pp 16648 - 16657. [2] www.jsnanotech.com [3] O’Shea, K.; Nash, R. An Introduction to Convolutional Neural Networks arXiv:1511.08458v2 (https://doi.org/10.48550/arXiv.1511.08458) [4] Agarap, A. F. Deep Learning using Rectified Linear Units (ReLU) arXiv:1803.08375v2 (https://doi.org/10.48550/arXiv.1803.08375) [5] Kingma, D. P.; Ba, J. L. Adam: A Method For Stochastic Optimization. 3rd International Conference on Learning Representations, ICLR 2015-Conference Track Proceedings, 2015; pp 1- 15.
