Automatic detection of lung nodules in CT scans has attracted significant research interest in computer-aided diagnosis (CAD) system because it can help improve the early and accurate diagnosis, therefore reducing lung cancer mortality. The aim of this paper is to propose a new method of lung nodule detection in CT scans by using 3D U-Net convolutional neural networks (CNNs) which can successfully reflect the 3D characteristics of CT scans. We use 3D U-Net architecture for CNNs in order to accurately and automatically detect the lung nodules in CT scans. We employ the masked Lung Image Database Consortium (LIDC) dataset containing 400000 CT images of over 1000 patients for training the 3D U-Net model. We introduce an integrated loss function in order to detect lung nodules with different sizes and prepare 3D input data by the use of 2D CT images fixing the hyper parameters of CNNs. Our method successfully detects the lung nodules with different sizes ranging from 3mm to 30 mm exhibiting higher sensitivity of 93.2, 94.5 and 94.8% at 1.0, 2.0 and 3.0 FPs per patient, respectively than the state-of-the-art methods. The proposed method of using 3D U-Net CNNs and integrated loss function is effective for early diagnosis of lung nodules in different sizes with improved sensitivity