A novel self-adaptive convolutional neural network model using spatial pyramid pooling for 3D lung nodule computer-aided diagnosis

This research proposes a novel self-adaptive convolutional neural network (Adap-Net) model for lung nodule diagnosis on 3D computed tomography (CT) images. Lung cancer is one of the most common cancers with a high mortality rate. Therefore, there is an urgent need to diagnose lung nodules to improve the survival rate, which is challenging because of the nodule heterogeneity and the lack of annotated lung nodule images. Prevailing research for lung nodule diagnosis usually ignores the nodule heterogeneity problem and enlarges the model complexity that degrades the lung nodule diagnosis performance given limited annotated training samples. To overcome the challenges, a transverse layer pooling (TLP) algorithm is proposed and the spatial pyramid pooling (SPP) scheme is integrated, which makes it possible to adaptively extract equal-dimensional feature representations from arbitrary-sized 3D lung nodule images. Meanwhile, the TLP algorithm introduces a layer compression architecture that dramatically reduces the model complexity. Moreover, K-means clustering is adopted to assign appropriate input image sizes for each lung nodule, allowing the mini-batch-based model training. The proposed Adap-Net is comprehensively evaluated and compared to other deep learning (DL) models using 3D CT images from a public dataset. Experimental results show that the proposed Adap-Net model improves the lung nodule diagnosis accuracy up to 12.12% with less than 10% of parameters that are involved in other DL models. In practice, the proposed Adap-Net model can offer complementary opinions in computer-aided diagnosis (CAD) systems as a supportive tool for radiologists and physicians in the medical image interpretation, analysis, and diagnosis process.