基于机器学习的医疗影像辅助诊断系统设计

Intelligent diagnosis of ophthalmic diseases based on machine learning

由于糖尿病视网膜病变(DR)、黄斑水肿、玻璃膜疣等疾病起病隐匿不易被察觉,导致误诊率、漏诊率较高,因此,本文基于深度学习模型的卷积神经网络,设计一款云端医疗诊断系统,辅助医生给患者进行诊断。在对这3种疾病的视网膜OCT图像分类时,因获取不到足够大的数据集,将数据增强、迁移学习等算法应用于OCT图像的识别中。在机器学习开发环境中训练模型后,将网络模型移植到ZYNQ平台上,以Xilinx推出的ZYNQ UltraScale+MPSoC作为处理器,创造性地将机器学习算法移植到嵌入式软硬件平台上,通过ARM与现场可编程逻辑门阵列(FPGA)的协同工作,达到机器学习常用的处理器GPU所不具备的性能。测试结果表明,经迁移学习后的GoogLeNet网络对OCT图像的识别正确率达98%左右,基本能达到一个经验丰富的医生水平。且本系统与GPU相比,具有处理速度快、功耗低、用户体验好、便携等特点,适合推广应用。

The three ophthalmic diseases of diabetic retinopathy(DR),macular edema,and drusen have the characteristics of insidious onset and are not easy to be detected,resulting in high rates of misdiagnosis and missed diagnosis by the medical imaging doctors in China.Based on the deep learning model of a convolutional neural network,this paper designs a cloud medical diagnosis system,and this system can assist doctors to make a diagnosis for patients.In the classification of retinal OCT images of these three diseases,due to the inability to obtain a large enough data set,apply data enhancement,transfer learning and other algorithms to recognize retinal OCT images.After training the model in the machine learning development environment,the network model is transplanted to the ZYNQ platform,and the ZYNQ UltraScale+MPSoC launched by Xilinx is used as a processor.The system creatively migrates the machine learning algorithms to embedded software and hardware platforms.Through the collaborative work of ARM and field programmable gate array(FPGA),the performance that is not available to the graphics processing unit(GPU)processor commonly used in machine learning is achieved.The results show that after the migration learning,the GoogLeNet network has an accurate recognition rate of about 98%for OCT images,which can basically reach the level of an experienced doctor.Compared with the GPU,the system has the characteristics of the fast processing speed,low power consumption,and good user experience and portability,which is suitable for popularization and application.