Segmentation of pulmonary nodules in CT images based on 3D‐UNET combined with three‐dimensional conditional random field optimization

Abstract

Purpose: Pulmonary nodules are a potential manifestation of lung cancer. In computer-aided diagnosis (CAD) of lung cancer, it is of great significance to extract the complete boundary of the pulmonary nodules in the computed tomography (CT) scans accurately. It can provide doctors with important information such as tumor size and density, which assist doctors in subsequent diagnosis and treatment. In addition to this, in the molecular subtype and radiomics of lung cancer, segmentation of lung nodules also plays a pivotal role. Existing methods are difficult to use only one model to simultaneously treat the boundaries of multiple types of lung nodules in CT images. Method: In order to solve the problem, this paper proposed a three-dimensional (3D)-UNET network model optimized by a 3D conditional random field (3D-CRF) to segment pulmonary nodules. On the basis of 3D-UNET, the 3D-CRF is used to optimize the sample output of the training set, so as to update the network weights in training process, reduce the model training time, and reduce the loss rate of the model. We selected 936 sets of pulmonary nodule data for the lung image database consortium and image database resource initiative (LIDC-IDRI) 1 database to train and test the model. What's more, we used clinical data from partner hospitals for additional validation. Results and conclusions: The results show that our method is accurate and effective. Particularly, it shows more significance for the optimization of the segmentation of adhesive pulmonary nodules (the juxta-pleural and juxta-vascular nodules) and ground glass pulmonary nodules (GGNs).

Automated detection and segmentation of non-small cell lung cancer computed tomography images

“…A 2D CNN architecture was chosen for several reasons: (1) by using a 2D input the training dataset can be increased by more than a factor of 60, as overall more than 60,000 unique slices were available in the training set; (2) due to calculation costs, most present deep 3D architectures could analyze only a subvolume of the medical image 31 , 32 , or they require a dimensionality reduction using interpolation or other image processing methods. 2D architectures do not have this problem and can process CT scans in the original resolution; (3) our main goal was to develop a pipeline that can be used in a clinical setting, and a 2D architecture allows for significantly lower requirements for executing PC.…”

Comparative study of deep neural networks with unsupervised Noise2Noise strategy for noise reduction of optical coherence tomography images

“…UNet [30], as a powerful neural network architecture, was first proposed for image segmentation and achieved excellent performance in many scenarios, such as lung nodules segmentation in CT image [37], prostate segmentation in MR image [38] and retinal layers segmentation in OCT images [39]. One of the most important superiorities of UNet is that the skip-connection part of different level is employed, which assembles the feature representations across multiple semantic scales and prevents information loss.…”

Explainable multiple abnormality classification of chest CT volumes with deep learning

“…To improve AxialNet, we propose a mask loss that is mathematically related to a typical loss function used for training an abnormality segmentation model. Prior work in abnormality segmentation in CT scans has focused on groundglass opacities and consolidation [32,33], COVID-19-related "anomalies" [34], pneumothorax [35], and lung nodules [36,37,38]. Our mask loss does rely on abnormality-specific allowed regions; however, unlike an abnormality segmentation loss, the mask loss is intended to enhance a classification model rather than train an abnormality segmentation model; it is calculated in a low-dimensional space for computational feasibility, rather than the input space; and it relies on automatically generated allowed regions rather than manually obtained abnormality segmentation maps, so that over 80 abnormalities across 36,316 CT volumes can be considered rather than 1-2 abnormalities across a few hundred CT volumes.…”