Intelligent And Portable Biological Cell Detection System Research Based On Lensless Microfluidic Imaging

With the development of smart medicine,telemedicine,and medical networking,there is a growing demand for immediate testing of biological cells and biological samples(POCT,Point of Care Testing).Lensless imaging systems with a large field of view and miniaturization have received widespread attention as sensing node devices for medical networking.However,the main function of current lensless cell detection systems is cell image acquisition.The algorithms for feature extraction and classification reconstruction of cell images still mainly rely on the upper computer or offline methods,which makes it difficult to realize the portability of detection devices.Around the application requirements of low-power,low-cost,and real-time terminal systems,this paper addresses the lack of cell image processing capability in portable terminals of lensless cell imaging systems,studies cell image processing based on neural networks,and deeply explores the hardware implementation scheme of the algorithms,with the main research results as follows:1.An asymmetric quantization algorithm for general neural network operations was proposed to solve the problem that convolution neural networks are computationally intensive and difficult to apply to portable devices.The algorithm dynamically adjusts the quantization centroid offset to achieve the best quantization effect according to the different distribution characteristics of the eigenvalue range and weight range of the convolutional neural network and reunifies the quantization parameters after the convolutional operation is completed.The experimental analysis shows that the recognition accuracy decreases only 0.56%when using int8 quantization compared with float32.The hardware circuit power consumption and area are effectively reduced while ensuring the accuracy of neural network operation.Then,the circuit architecture of the quantization algorithm is further constructed,and a circuit design scheme of dual-configuration register sets is proposed to effectively improve the computational efficiency between the neural network operation layers.The circuit of asymmetric integer low-bit quantization algorithm is implemented on FPGA and applied to the CNN leukocyte classification system.The results show that the int8 quantization can obtain a classification speed of 17.9fps with 98.44%accuracy when the FPGA frequency is 100MHz.Compared with float32,the circuit area decreases by 45%and the recognition accuracy decreases by only 0.56%.2.The complementary algorithm based on the Booth multiplier,and the Winograd algorithm optimized for hardware implementation were proposed to address the problem of high energy consumption of the neural network acceleration circuit,after making an exploratory study of the convolutional array circuit.The complementary algorithm based on the Booth multiplier removes the subtractors from the LUT operation using uniform compensation of the complement to achieve the purpose of reducing circuit complexity and power consumption,while the Winograd algorithm optimized for hardware implementation removes the subtractor design from the circuit by counting the subtractors during the entire Winograd operation and compensates all subtractors uniformly at the end of the operation using complementary compensation.The consumption of circuit area and power consumption during the operation of convolutional neural networks are effectively reduced by these two algorithms.The algorithm circuit clock domain divided by multi-level operation accuracy level,and the control method for dynamically adjusting the accuracy of neural network operation by power level are proposed,when the algorithm acceleration circuit architecture is further constructed.The control method can work at different operation precision based on different battery levels,thus effectively improving the operating time of battery-powered detection devices.The complement optimization algorithm based on the Booth multiplier,and the Winograd algorithm optimized for hardware implementation are implemented on FPGAs and applied to the yeast culture detection system and algae detection system based on lensless imaging,respectively.Experiments show that the complementary algorithm circuit based on the Booth multiplier can reduce the area by 18.11%and power consumption by 23.5%;the Winograd algorithm circuit optimized for hardware implementation has a 55.56%higher operation speed,3.19%lower circuit area,and 11.09%lower power consumption compared with the ordinary convolutional circuit;Compared with the fixedprecision method,the dynamic neural network precision adjustment control method based on the battery power of the device can increase the operating time of the device by 101.04%and decrease the recognition precision by only 6.47%under the same battery power condition.3.A smart internal of things(IoT)architecture with a three-level structure of cloud-sideterminal was proposed for the interconnected information transfer needs of portable biological cell detection devices on the smart IoT after the prototype system for acquisition and analysis is complete.The intelligent portable biological cell detection device proposed in this paper is used as a side and terminal device of the intelligent IoT to complete the collection and wireless transmission of multiple biological cell information and complete the preliminary intelligent analysis of the data,and finally transmit the preliminary analysis results of the data to the cloud.This work explores the design method that can integrate multiple functions of biological cell image acquisition and cultivation into a small multi-functional image acquisition device,and gives the design method of AIoT multi-level intelligent detection network for wireless Bluetooth multi-channel acquisition.4.For different cell culture and detection needs,microfluidic channel chips and image acquisition hardware adapted to the characteristics of different application needs were studied,designed,and fabricated,and different portable cell culture and monitoring systems were finally designed and completed.At the same time,the temperature control of biological cell detection devices was studied and explored in these systems,and a design method that can realize an automatic temperature control system on portable biomonitoring devices was proposed.These systems include various intelligent biological cell detection systems such as statistical analysis of blood cell classification,algae monitor,algae culture,yeast culture,etc.In summary,the feasibility and research value of the portable intelligent cell culture detection system is based on microfluidic chip lens-free microscopic imaging,which provides the possibility of combining microelectronics,artificial intelligence,and cell detection,and provides a direction for miniaturization of biological cell detection.It is beneficial to promote the transformation of artificial intelligence(AI)cell culture collection and analysis from large dedicated servers to portable cell analysis devices and to promote rapid early analysis of major diseases and environmental pollution.