In China, lung cancer ranks first in both incidence and mortality rates, accounting for 17.9% of all new cases and 23.8% of total cancer deaths according to GLOBOCAN 2020 (1). Non-small-cell lung cancer (NSCLC) constitutes the majority of lung cancers. Radiotherapy (RT) is usually used in all stages of NSCLC treatment and is required at least once in more than half of patients for either cure or palliation. In a typical clinical workflow of RT, a radiation oncologist manually segments the tumor target and organs at risk (OARs) based on the information provided by CT, MRI and/or PET/CT images (2, 3). This process is often time consuming and laborious, and the quality of the segmentations largely depends on the experience of radiation oncologists. It is easy to distinguish the organs with high contrast on CT images; however, it is difficult to distinguish the boundary between tumor tissue and surrounding normal tissue with similar contrast. Moreover, inconsistencies in target and OARs segmentations have been reported for both inter-and intraobserver segmentation variability (4–8). These factors will affect the accuracy and efficacy of RT. Therefore, improving the consistency and efficiency of image segmentation becomes an urgent task.

In recent years, automatic medical image segmentation based on deep learning has become a popular research topic in RT, and several convolutional neural networks (CNNs) including U-Net, ResNet and DenseNet, have shown great success in autosegmentation of the target and OARs (9–16). DenseNet was proposed by Huang G et al. (17) in 2017, using the concept of feature map reuse to address the small training datasets in supervised learning. Moreover, DenseNet connects multiple dense blocks with a transition layer and concatenates the channels of each dense block feature map in series to increase the number of feature maps and improve the utilization rate of feature maps. Tong N et al. (18) improved the performance of their previous shape constrained fully CNNs for head and neck OARs segmentation on CT and low field MRI by incorporating generative adversarial network (GAN) and DenseNet. With the novel segmentation method, they showed that the low field MR images acquired on a MR guided radiation radiotherapy system can support accurate and fully automated segmentation of both bony and soft tissue OARs for adaptive radiotherapy. Fu J et al. (19) proposed a novel three-dimensional (3D) multipath DenseNet for generating the accurate glioblastoma (GBM) tumor contour from four multimodal pre-operative MR images. The multipath DenseNet demonstrated an improved accuracy over comparable algorithms in the clinical task of GBM tumor segmentation. To our best knowledge, there has not been an automatic segmentation study based on the DenseNet for NSCLC radiotherapy.

In this study, a deep learning model based on DenseNet and FCN (fully convolutional network) is proposed. The model uses the idea of feature reuse. It learns the planar distribution characteristics of OARs in CT images through a denseblock module and supplements details through long connections to achieve an end-to-end accurate OAR delineation for NSCLC patients.

The results of this study are relatively consistent with those of the challenge report of automatic segmentation of thoracic organs organized by the American Association of Physicists in Medicine (AAPM)’s annual meeting in 2017 (21),with the right lung having the highest average DSC (0.96) and the esophagus having the lowest average DSC (0.67). Compared with U-Net, the autosegmentation results of the OARs based on the proposed model were better with higher DSC as well as lower HD95 and ASD. Among them, DSC differences of the spinal cord and esophagus, HD95 differences of the spinal cord, heart, trachea and esophagus, as well as ASD difference of the spinal cord were statistically significant (P<0.05).

Lustberg T et al. (23) used a deep learning autosegmentation software (Mirada) to create thoracic OARs contours and the model was built by using 450 lung patients’ images as the training set. For 20 CT scans of stage I-III NSCLC patients in the testing set, the median DSCs of the spinal cord, the lungs, and heart were 0.83, >0.95, >0.90, respectively. Zhang T et al. (24) developed a 2D AS-CNN based on the ResNet101 network using a training dataset of 200 lung cancer patients. The average DSCs of the left lung, right lung, heart, spinal cord, and esophagus of 19 NSCLC patients were 0.94, 0.94, 0.89, 0.82, and 0.73, respectively. Zhu JH et al. (25) proposed an automatic segmentation model based on depth convolution to segment CT images from 36 lung cancer patients. The average DSCs of the lungs, heart, liver, spinal cord and esophagus were 0.95, 0.91, 0.89, 0.76 and 0.64, respectively. Dong X et al. proposed a U-Net-generative adversarial network (U-Net-GAN) and realized the segmentation of 5 thoracic OARs. Among them, the left lung, right lung, and heart were automatically segmented by a 2.5D GAN model, while the esophagus and spinal cord were automatically segmented by a 3D GAN model. The average DSCs of the left and right lungs, spinal cord, esophagus, and heart were 0.97,0.97, 0.90, 0.75, and o.87, respectively. He T et al. (26) proposed a uniform U-like encoder-decoder architecture abstracted from the U-Net and trained it using 40 patients’ thoracic CT scans. High DSC values were obtained for esophagus (0.86), heart (0.95), trachea(0.92) and aorta (0.95) from 20 patients in the testing set. Feng X et al. (27) developed a model based on 3D U-Net to autosegment thoracic OARs using 36 thoracic CT scans as the training set. The performance of the model was evaluated on two groups of testing set consisting of 12 patients and 30 patients, respectively. The average DSCs of the spinal cord, right lung, left lung, heart and esophagus of the first testing set reached 0.89, 0.97, 0.98, 0.93, and 0.73 while those of the second testing set were 0.85, 0.98, 0.98, 0.86 and 0.69, respectively.

The differences in all dosimetric metrics of the OARs between manual and automatic delineations were not statistically significant (P>0.05) in our study. The maximum dosimetric metrics differences were 0.41Gy for Dmean of the trachea and 0.64% for V5 of bilateral lungs, while the clinically acceptable dose difference and volume difference of OARs between manual and automatic delineation are supposed to be within 1Gy and 1%, respectively. Zhu J et al. (28) evaluated the performance of automatic segmentation of the OARs with dosimetric metrics for esophageal cancer patients. The maximum metrics differences were 0.35 Gy for Dmax of the spinal cord and 0.4% for V30 of bilateral lungs. The results in our study were consistent with those of the above study.

Due to the different training datasets, it is difficult to compare the advantages and disadvantages of the proposed model and the published model. However, the number of training cases used in our study was obviously fewer; the proposed model has strong feature extraction ability in the training of small samples, and the segmentation results are similar to those of the training model of relatively large datasets. A limitation of this study needs to be pointed out. That is, due to low soft tissue contrast, small volume, and large shape variability across patients, the automatic segmentation results of the esophagus are not ideal, and the DSC value is lower than 0.7, which is clinically unacceptable (29, 30), therefore, we did not take into account the esophagus when analyzing the dose-volume parameters in the treatment plan. In the next work, we need to further optimize the model and expand the size of data to increase its generalization and segmentation effect.

Currently, there are three main development directions for deep learning networks in medical image segmentation. The first direction is to deepen the network level and depth, extract deeper semantic features to obtain stronger expression ability, or widen the network to increase the number of channels to obtain more information in the same layer, such as the texture features of different grayscales and boundary features in different directions. The second direction is to obtain a more effective spatial feature extraction ability by learning the sequence concatenation properties of multiple CT slices of a patient, represented by 3D U-Net and many other derivative networks. The third direction represented by DenseNet is to improve the utilization rate of the feature map by sharing the feature map layer by layer to enhance the feature expression ability of the image and improve the generalization performance of the network (31).

Compared with U-Net, the proposed model based on DenseNet is better in the OARs segmentation task; even if the training set has fewer images, it can still fairly effectively prevent the occurrence of overfitting. At the same time, it can effectively alleviate the problem of the gradient disappearing in the training process by repeatedly using different levels of feature maps, providing a new idea for medical image segmentation.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by the Ethics Committee of the Seventh Medical Center of Chinese PLA General Hospital. The patients/participants provided their written informed consent to participate in this study.

FZ and QW contributed conception and design of the study. FZ and AY trained the deep learning models. FZ and QW performed data analysis and drafted the manuscript. NL, HJ, DC, and YW helped to collect the data and evaluate radiotherapy planning. YY designed radiotherapy planning. All authors contributed to the article and approved the submitted version.

This work was supported by the Beijing Municipal Science and Technology Commission (No.Z181100001718011).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.