Thomas Johann Seebeck Department of Electronics, Tallinn University of Technology, Tallinn, 19086, Estonia.
Eliko Tehnoloogia Arenduskeskus OÜ, Tallinn 12618 and OÜ, Tallinn, 10143, Estonia.
Med Phys. 2018 Mar;45(3):1135-1149. doi: 10.1002/mp.12746. Epub 2018 Jan 23.
Currently reported computer-aided detection (CAD) approaches face difficulties in identifying the diverse pulmonary nodules in thoracic computed tomography (CT) images, especially in heterogeneous datasets. We present a novel CAD system specifically designed to identify multisize nodule candidates in multiple heterogeneous datasets.
The proposed CAD scheme is divided into two phases: primary phase and final phase. The primary phase started with the lung segmentation algorithm and the segmented lungs were further refined using morphological closing process to include the pleural nodules. Next, we empirically formulated three subalgorithms modules to detect different sizes of nodule candidates (≥3 and <6 mm; ≥6 and <10 mm; and ≥10 mm). Each subalgorithm module included a multistage flow of rule-based thresholding and morphological processes. In the final phase, the nodule candidates were augmented to boost the performance of the classifier. The CAD system was trained using a total number of nodule candidates = 201,654 (after augmentation) and nonnodule candidates = 731,486. A rich set of 515 features based on cluster, texture, and voxel-based intensity features were utilized to train a neural network classifier. The proposed method was trained on 899 scans from the Lung Image Database Consortium/Image Database Resource Initiative (LIDC-IDRI). The CAD system was also independently tested on 153 CT scans taken from the AAPM-SPIE-LungX Dataset and two subsets from the Early Lung Cancer Action Project (ELCAP and PCF).
For the LIDC-IDRI training set, the proposed CAD scheme yielded an overall sensitivity of 85.6% (1189/1390) and 83.5% (1161/1390) at 8 FP/scan and 1 FP/scan, respectively. For the three independent test sets, the CAD system achieved an average sensitivity of 68.4% at 8 FP/scan.
The authors conclude that the proposed CAD system can identify dissimilar nodule candidates in the multiple heterogeneous datasets. It could be considered as a useful tool to support radiologists during screening trials.
目前报道的计算机辅助检测（CAD）方法在识别胸部计算机断层扫描（CT）图像中的各种肺结节方面存在困难，特别是在异质数据集。我们提出了一种新的 CAD 系统，专门用于识别多个异质数据集中的多尺寸结节候选物。
所提出的 CAD 方案分为两个阶段：主要阶段和最终阶段。主要阶段从肺部分割算法开始，然后使用形态学闭合过程进一步细化分割后的肺部，以包括胸膜结节。接下来，我们经验性地制定了三个子算法模块来检测不同大小的结节候选物（≥3 且<6mm；≥6 且<10mm；以及≥10mm）。每个子算法模块包括基于规则的阈值和形态学处理的多阶段流。在最终阶段，结节候选物被扩充以提高分类器的性能。CAD 系统使用总共 201654 个结节候选物（扩充后）和 731486 个非结节候选物进行训练。利用基于聚类、纹理和体素强度特征的丰富的 515 个特征集来训练神经网络分类器。该方法在 Lung Image Database Consortium/Image Database Resource Initiative（LIDC-IDRI）的 899 个扫描中进行了训练。CAD 系统还独立地在 AAPM-SPIE-LungX 数据集的 153 个 CT 扫描和 Early Lung Cancer Action Project（ELCAP 和 PCF）的两个子集中进行了测试。
对于 LIDC-IDRI 训练集，所提出的 CAD 方案在 8 FP/scan 和 1 FP/scan 时分别产生了 85.6%（1189/1390）和 83.5%（1161/1390）的总体灵敏度。对于三个独立的测试集，CAD 系统在 8 FP/scan 时平均灵敏度为 68.4%。
作者得出结论，所提出的 CAD 系统可以识别多个异质数据集中的不同结节候选物。它可以被认为是支持筛查试验中放射科医生的有用工具。
