【摘要】 目的为实现对脑部数字减影血管造影(DSA)序列的自动化判读,探索在DSA影像中目标检测算法FasterRCNN对复杂血管结构鉴别的应用性。方法收集来自复旦大学附属华山医院DSA影像库2010年1月至2013年12月的正常颈内动脉正位造影图像共计388例,其中350例作为模型训练测试集(测试集)数据,38例作为独立验证集。①测试集DSA中筛选出曝光适度、显影清晰的影像共计680张,比例为8∶2。根据不同时期DSA影像的血管特征,标记不同的感兴趣区域,图片集总计标注了5类血管特征区域。搭建Faster-RCNN多目标检测网络,优化网络参数,保存最优模型。分析测试集各类血管结构的平均精度(AP)和多类别平均精度均值(mAP)。②独立验证集DSA数据依次输入模型进行血管结构检测,分析各图像血管结构的类别与出现的时间,以此为标准对每张图像的时相进行区分,从而确定每一例DSA的时相区间。将判定结果与专科医生标定的结果进行比较,计算各时期的区分准确率。结果测试集136张图片中,颈内动脉的AP为0.922、Willis环为0.991、大静脉为0.899、静脉血管为0.769、静脉窦为0.929。5类血管特征区域的多类别mAP为0.902。独立验证集中,动脉期、毛细血管期、静脉早期和静脉窦期分期准确率分别达到100%,92.1%,92.1%和78.9%。结论 Faster-RCNN算法可以分析DSA序列中的时间信息与结构信息从而对DSA影像进行自动判读,可在缩短读片时间前提下保证足够的判读准确度,为复杂脑血管的鉴别提供技术支持。更多还原

【Abstract】 Aim In order to realize the automatic reading of brain digital subtraction angiography(DSA) sequences, and to explore the applicability of the target detection algorithm FasterRCNN(faster region-based convolutional networks) in the identification of complex vascular structures in DSA images. Methods A total of 388 normal internal carotid arterial angiography images from the DSA image library of Huashan Hospital affiliated to Fudan University from January 2010 to December 2013 were collected, of which 350 were used as the model training-test set(test set) data, and 38 were used as independent verification set. ① In the test set DSA, a total of 680 images with moderate exposure and clear development were screened out, with a ratio of 8:2. According to DSA images in different phase, different regions of interest are marked, and the picture set has a total of 5 types of vessel regions. A Faster-RCNN multi-target detection network was built, network parameters were optimized, and the optimal model was saved. The average precision(AP) and mean average precision(mAP) of vascular structures in the test set were analyzed. ② The DSA data of the independent verification set were sequentially inputted into the model to detect the blood vessel structure, analyze the type and time of appearance of the blood vessel structure in each image, and use it as the standard to distinguish the time phase of each image and then determine the time phase interval of each case of DSA. After comparing the judgment results with the calibration results of the specialist, the discrimination accuracy rate can be calculated in each period. Results Among the 136 images in the test set, the AP of the internal carotid artery was 0.922, the circle of Willis was 0.991, the large vein was 0.899, the venous vessel was 0.769, and the venous sinus was 0.929. The mAP of the five types of vascular was 0.902. In the independent verification, the accuracy in the arterial phase, capillary phase, early venous phase and venous sinus phase reached 100%, 92.1%, 92.1% and 78.9%, respectively. Conclusion Faster-RCNN algorithm can be used to analyze the time information and structural information in the DSA sequence, and automatically read the DSA images, the sufficient interpretation accuracy under the premise of shortening the reading time can be ensured, and provide technical support for the identification of complex cerebral vessels.更多还原