基于深度学习的三叉神经区域自动检测及TensorRT加速

Automatic detection of trigeminal neural region based on deep learning and TensorRT acceleration

摘要:

利用深度学习技术对颅脑核磁共振图像(MRI)中三叉神经区域进行自动检测可为后续三叉神经分割提供可靠的输入图像，有效解决了人工筛选三叉神经对临床医生专业素养要求高、耗时长等弊端。采用YOLO网络自动检测颅脑核磁共振图像中三叉神经区域提高推理速度，并系统性地评估NVIDIA TensorRT框架在不同计算平台下的推理性能。实验结果表明，通过YOLO目标检测网络能够准确检测出三叉神经所在的区域，同时在NVIDIA TensorRT框架下，当输入的颅脑MRI分辨率为(204×204)时，CPU平台、嵌入式GPU平台、桌面GPU平台及专业GPU计算卡平台下，YOLOv2网络检测优化后的三叉神经目标的每秒帧率分别可达到0.1 FPS,23.4 FPS,112.5FPS和793.7 FPS，这为后续开发便携式的三叉神经分割设备提供了可参考的重要依据。

Abstract:

Manual screening of trigeminal nerves requires high professional quality and is time consuming for clinicians. Using deep learning to automatically detect trigeminal nerve regions in cranial Magnetic Resonance Imaging (MRI) can provide a reliable input image for subsequent trigeminal nerve segmentation. YOLO(You Only Look Once) network is utilized to automatically detect the trigeminal nerve region of the cranial magnetic resonance image to improve the inference speed, and to systematically evaluate the inference performance of the NVIDIA TensorRT framework under different computing platforms. The experimental results show that the YOLO target detection network can accurately detect the area where the trigeminal nerve is located. Simultaneously, under the NVIDIA TensorRT framework, when the input brain MRI resolution is (204×204), the YOLOv2 network detects the optimized trigeminal nerve through the CPU platform, embedded GPU platform, desktop GPU platform and professional GPU computing card platform, the frame rates per second can reach 0.1 FPS, 23.4 FPS, and 793.7 FPS . This provides important reference for the subsequent development of portable trigeminal neural segmentation equipment.