基于尺寸自适应深度神经网络的胸部CT图像肺结节检测

Size-Adaptive Deep Neural Networks Based Pulmonary Nodule Detection in CT Scans

使用计算机断层扫描(CT)筛查肺结节是早期肺癌诊断的重要手段。但由于肺结节在形状、大小和位置上有存在很大的差异,目前肺结节尤其是小结节的自动检测依然具有挑战性。为了实现高灵敏度的肺结节检测,提出一种新的计算机辅助检测系统,该系统采用两种新的策略:尺寸自适应候选检测(SACD)和尺寸自适应假阳性抑制(SAFPR)。首先,SACD结合深层和浅层卷积特征构建高级特征,以获得感兴趣区域的位置和大小信息;然后将检测结果送入用于筛查不同大小结节的3个并行子网络,从而细化SACD的检测结果,提高检测系统的准确性和鲁棒性。在LIDC-IDRI数据集(1 186个结节)上的验证结果表明,该系统在1 FPs/扫描的情况下,可以达到96%的灵敏度,其性能可达到或优于最新的智能诊断系统。在包含430个结节的独立数据集实验结果表明,对于(3.89±2.34) mm的结节,在0.3FPs/扫描时,该系统的检测灵敏度为69.53%,与两位经验丰富的放射科医生的人工筛查结果相当,表明该系统具有一定的临床应用价值。

Computed tomography( CT) screening for pulmonary nodules is an important method for early diagnosis of lung cancer. However,the automatic detection of pulmonary nodules,especially small nodules,is still challenging due to the large differences in shape,size and location of pulmonary nodules. To achieve highly sensitive detection of pulmonary nodules,in this paper,a new computer-aided detection system for pulmonary nodules detection was proposed. The system adopted two new strategies: size adaptive candidate test( SACD)and size adaptive false positive reduction( SAFPR). First,SACD combined deep and shallow convolution features to construct advanced features and detected CT images to obtain the location and size information of the region of interest. Then,the detection results were sent to three parallel sub-networks for screening different sizes of nodules,so as to refine the detection results of SACD and improved the accuracy and robustness of the computer-aided detection system. The results on the LIDC-IDRI dataset( 1186 nodules) demonstrated that the proposed system achieved the high sensitivity of 96% at 1 FPs/scan,which was superior or comparable to the state-of-the-art systems,while in an independent dataset containing 430 nodules,the detection sensitivity of the system was 69. 53% at 0. 3 FPs/scan for the nodules with the size of 3. 89±2. 34 mm,which was comparable to the human screening results of two experienced radiologists,indicating that the system has certain clinical application value.