CDAE-C: A Fully Convolutional Denoising Auto-Encoder with 2.5D Convolutional Classifier

Lung cancer, a major burden of disease causing most cancer related deaths worldwide, can be well treated with early diagnosis of malignant lung nodule using high resolution chest computed tomography. Low dose computed tomography, with lower radiation risk to patients than normal dose computed tomography, benefits the health of patients but degrades the image quality with interfering noise, which can compromise diagnostic performance. In this paper, a multi-function model is introduced that deals with both lung nodule classification and CT image noise deduction. The proposed model consists of a fully convolutional denoising auto-encoder and a 2.5D convolutional classifier and is referred as Convolutional Denoising Auto-Encoder and Classifier (CDAE-C). The training of the proposed model is conducted following a two-phase process in which CDAE is firstly trained to denoise and reconstruct low-dose CT images and then CDAE-C is trained on latent code from pretrained encoder and 3D spatial relationship of lung nodules to classify benign and malignant lung nodules. Fully convolutional structure of denoising auto-encoder ensures the model can accept and reconstruct a low-dose CT image independent of its size, which is practical and very beneficial to the 2.5D classifier as the classification work of benign and malignant lung nodules needs regions of interest cropped from whole low-dose CT images. Extracting lung nodule’s latent representation from the pretrained encoder and using 3D spatial relationship of cropped lung nodule slices, 3D embedded features of each lung nodule are constructed as input of the proposed 2.5D convolutional classifier. Experimental results indicate that CDAE’s denoising performance is of RMSE\approx0.0458 and PSNR\approx27.2004, and CDAE-C classification performance reaches recall\ rate\approx97.67%, AUC\ \approx99.45% and FNR\approx2.17%. After ablation experiment, the proposed model is proved to have higher accuracy and convergence speed.