Ensemble Learning of Multiple-View 3D-CNNs Model for Micro-Nodules Identification in CT Images

Numerous automatic systems of pulmonary nodules detection have been proposed, but very few of them have been consecrated to micro-nodules (diameter < 3 mm) even though they are regarded as the earliest manifestations of lung cancer. Moreover, most available systems present high false positive rate resulting from their incapability of discriminating between micro-nodules and non-nodules. Thus, this paper proposes a system to differentiate between the micro-nodules and non-nodules in computed tomography (CT) images by an ensemble learning of multiple-view 3-D convolutional neural networks (3D-CNNs). A total of 34 494 volumetric image samples, including 13 179 micro-nodules and 21 315 non-nodules, are acquired from the 1010 CT scans of the LIDC/IDRI database. The pulmonary nodule candidates are cropped with five different sizes, including
$20\times 20\times 3$
,
$16\times 16\times 3$
,
$12\times 12\times 3$
,
$8\times 8\times 3$
, and
$4\times 4\times 3$
. Then, five distinct 3D-CNN models are built and implemented on one size of the nodule candidates. An extreme learning machine (ELM) network is utilized to integrate the five 3D-CNN outputs, yielding the final classification results. The performance of the proposed system is assessed in terms of accuracy, area under the curve (AUC), F-score, and sensitivity. It is found that the proposed system yields an accuracy, AUC, F-score, and sensitivity of 97.35%, 0.98, 96.42%, and 96.57%, respectively. These performances are highly superior to those of 2D-CNNs, single 3D-CNN model, as well as those by the state-of-the-art methods implemented on the same dataset. For the ensemble method, ELM achieves better performance than the majority voting, averaging, AND operator, and autoencoder. The results demonstrate that developing an automatic system for discriminating between micro-nodules and non-nodules in CT images is feasible, which extends lung cancer studies to micro-nodules. The combination of multiple-view 3D-CNNs and ensemble learning contribute to excellent identification performance, and this strategy may help develop other reliable clinical-decision support systems.