Accepted Article
Article Type: Research Article Research Article
Automatic segmentation of the clinical target volume and organs at risk in the planning CT for rectal cancer using deep dilated convolutional neural networks
1
Kuo Men1, Jianrong Dai1# and Yexiong Li1#
National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking
Union Medical College, Beijing, 100021, China
#
Corresponding authors
Jianrong Dai Email:
[email protected] Yexiong Li Email:
[email protected] Byline: Segmentation of CTV and OARs using DDCNNs
This article has been accepted for publication and undergone full peer review but has not been through the copyediting, typesetting, pagination and proofreading process, which may lead to differences between this version and the Version of Record. Please cite this article as doi: 10.1002/mp.12602 This article is protected by copyright. All rights reserved.
Abstract
Accepted Article
Purpose: Delineation of the clinical target volume (CTV) and organs at risk (OARs) is very important for radiotherapy but is time-consuming and prone to inter-observer variation. Here, we proposed a novel deep dilated convolutional neural network (DDCNN)-based method for fast and consistent auto-segmentation of these structures.
Methods: Our DDCNN method was an end-to-end architecture enabling fast training and
testing. Specifically, it employed a novel multiple-scale convolutional architecture to extract
multiple-scale context features in the early layers, which contain the original information on fine texture and boundaries and which are very useful for accurate auto-segmentation. In addition, it enlarged the receptive fields of dilated convolutions at the end of networks to capture complementary context features. Then, it replaced the fully connected layers with fully convolutional layers to achieve pixel-wise segmentation. We used data from 278 patients with rectal cancer for evaluation. The CTV and OARs were delineated and validated by senior radiation oncologists in the planning computed tomography (CT) images. A total of 218 patients chosen randomly were used for training, and the remaining 60 for validation. The Dice similarity coefficient (DSC) was used to measure segmentation accuracy.
Results: Performance was evaluated on segmentation of the CTV and OARs. In addition, the performance of DDCNN was compared with that of U-Net. The proposed DDCNN method outperformed the U-Net for all segmentations, and the average DSC value of DDCNN was
3.8% higher than that of U-Net. Mean DSC values of DDCNN were 87.7% for the CTV, 93.4% for the bladder, 92.1% for the left femoral head, 92.3% for the right femoral head, 65.3% for the intestine and 61.8% for the colon. The test time was 45 s per patient for segmentation of all the CTV, bladder, left and right femoral heads, colon and intestine. We also assessed our approaches and results with those in the literature: our system showed superior performance and faster speed.
This article is protected by copyright. All rights reserved.
Conclusions: These data suggest that DDCNN can be used to segment the CTV and OARs
Accepted Article
accurately and efficiently. It was invariant to the body size, body shape, and age of the patients. DDCNN could improve the consistency of contouring and streamline radiotherapy
workflows.
Keywords: automatic segmentation, clinical target volume, organs at risk, deep learning, deep dilated convolutional neural networks, radiotherapy
1. INTRODUCTION As an effective treatment modality for cancer, radiotherapy technologies provide steep dose gradients to facilitate dose escalation to tumor targets while sparing organs at risk (OARs). Computed tomography (CT) is used routinely for simulation to generate volumetric information of the body due to its convenience and ability to provide the relative electron density1. Accurate delineation of the clinical target volume (CTV) and OARs on CT-simulation images is required to elicit therapeutic advantages and reduce the risk of irradiation of normal tissues. This task is also referred to as “image segmentation” and is usually carried out by radiation oncologists based on recommended guidelines. However, large inter- and intra-observer variation in delineation of the regions of interest (ROIs) has been reported2–5 due to the experience and preference of the operator. Also, such “manual” delineation is difficult and time-consuming.
Computer-assisted automatic segmentation of images can relieve radiation oncologists from the labor-intensive aspects of their work as well as increase the accuracy, consistency, and reproducibility of ROI delineation. In general, image-segmentation approaches can be classified into three methods based on region, edge, and classification6. These methods are
used to extract features from the intensity, gradient and texture, respectively. However, due to
This article is protected by copyright. All rights reserved.
the low contrast to noise ratio (CNR) and high-density artifacts of CT images, methods based
Accepted Article
on gray-level information make segmentation of ROIs with boundary insufficiencies very challenging. A priori knowledge incorporated into the process of segmentation may improve accuracy. In recent years, deep learning methods have been applied to various problems, especially in the field of computer vision7–14. They have superior performance than the previous state-of-the-art methods employed in many of these applications. Instead of extracting and learning informative features that describe the patterns manually, deep learning methods discover the informative representations in a self-taught manner and use hierarchical layers of learned abstraction to accomplish high-level tasks efficiently. Various types of deep learning algorithms are used in research, such as stacked auto-encoders (SAEs)15, deep belief
networks (DBNs)16, restricted Boltzmann machines (RBMs)17, recurrent neural networks
(RNNs)18, and convolutional neural networks (CNNs)19. The success and convenience of deep learning and the advent of digital imaging have prompted researchers to investigate applications in the analyses of medical images. However, in the deep models of SAEs, DBNs, and DBMs, the inputs are always in vector form, which results in loss of the structural information among neighboring pixels that is very important for medical images. This loss will inevitably destroy such information in images. CNNs take two-dimensional (2D) or 3D images as input that can utilize spatial and structure information effectively and thus become the main deep learning methods for analyses of medical images. In its current iteration, CNNs are composed of three types of layers: convolutional, pooling, and fully-connected19. One typical arrangement is alternation between convolutional and pooling layers, followed by fully-connected hidden layers. Unlike fully connected neural networks, CNNs exploit three mechanisms (a local receptive field, weight sharing, and subsampling) that drastically reduce the amount of parameters that must be learned in a model. There are different types of architectures of CNNs, such as LeNet19, AlexNet7, ZFNet20, VGGNet8, GoogLeNet21, ResNet22, and faster R-CNN11. In the analyses of medical images, CNNs have been applied successfully to classification23–25, detection26–28, segmentation29–51, registration52,53, and image generation/enhancement54–57.
This article is protected by copyright. All rights reserved.
Segmentation of medical-image is usually the first step in disease detection and treatment. It
Accepted Article
typically identifies the set of voxels making up an ROI, which can be realized by applying various deep learning methods (including CNNs and RNNs) in medical imaging. Many related studies have been reported in different segmentation tasks, such as cells31, nuclei32,
blood vessels33,34, optic disks34, neuronal structures35, brain tissue36, brain lesion37,38, hippocampus39, striatum40, ventricles41,42, liver43, kidneys44, pancreas45, prostate gland46,
bladder47, colon48, vertebrae49, placenta50 and muscle51. In spite of the good performance of deep learning methods on these segmentation tasks, the studies are confined mostly to the
field of radiology.
For the auto-segmentation of CTV and OARs in radiotherapy, only “Atlas-based auto-segmentation (ABAS)”58–63 methods are available in clinical software. However, ABAS
has two main limitations. First, a “universal Atlas” is difficult to build because of the inconsistencies in organs in patients of different ages, shapes, and sizes. Second, it is
time-consuming due to the deformable registration process. There has been increasing interest in segmentation of CTV and/or OARs for radiotherapy with machine learning
methods but few attempts have been presented. Qazi et al.64 segmented OARs in CT images
of the head and neck with a feature-driven model-based approach. Gomathi et al.65 proposed
a hybrid medoid shift with K-means (HMSK) algorithm for segmentation on CT images. Then, they segmented the spliced organ from the initial segmentation results to avoid
inappropriate segmentation66. Dolz et al.67 used support vector machines (SVM) successfully
to segment the brainstem on magnetic resonance imaging (MRI) in the context of brain cancer. Recently, while we were preparing this manuscript, Ibragimov et al.68 published a study using typical CNNs for OARs segmentation in CT images of the head and neck (which was the first report on automatic delineation with CNNs in radiotherapy), and Dolz et al.69 segmented OARs on the optic region using the stacked denoised auto-encoders (SDAEs) method. However, reports focusing on CTV segmentation using any deep learning methods are lacking. Unlike OARs, the CTV is not a region with clear boundaries but instead includes tissues of potential tumor spread or subclinical diseases that are barely detectable in planning This article is protected by copyright. All rights reserved.
CT images. Segmentation of the CTV depends largely on the physician’s knowledge. Deep
Accepted Article
learning could relieve the use of domain expert knowledge in the design and extraction of the most appropriate discriminative features.
Here, we first present a novel deep dilated convolutional neural network (DDCNN) architecture for segmentation of the CTV and OARs in the planning CT for rectal cancer. Based on popular CNN architectures, we replaced the fully connected layers with fully convolutional layers to achieve pixel-wise segmentation in CT images. To solve the inconsistencies from person to person, a dilated convolution filter was introduced to achieve multiple-scale and multiple-level feature learning. The dilated convolution filter augments the size of the sampling space, which allows CNNs to extract features in larger receptive fields. Different from traditional CNNs architectures that employ dilated convolution at the higher
or middle layer in the convolutional network, the proposed DDCNN not only used the dilated convolution at the back of the network layer, but also directly deployed a dilated convolution module on the input image to capture multi-scale context information. These contextual
features can capture the finer structural information about the intensity, texture and contour that are essential for accurate localization. We also compared its performance with that of U-Net29, which is used widely as a baseline for segmentation tasks in medical images.
This study was novel with four main contributions. First, there were no reports on segmentation of the CTV in planning CT images for radiotherapy with any CNNs. Furthermore, this work was also the first attempt to apply CNNs for auto-segmentation of the CTV and OARs in planning of radiotherapy of rectal cancer. Second, a contrast-limited adaptive histogram equalization (CLAHE) algorithm was used to pre-process the input data for image enhancement. It could improve local contrast and enhance the definitions without over-amplifying the noise. Third, a multiple-path dilated convolution was introduced to extract original context information from input images directly, and compensate contextual features into high-level convolutional layers and thus improve the segmentation accuracy
This article is protected by copyright. All rights reserved.
compared with that for typical CNNs68. Finally, the proposed method was invariant to the
Accepted Article
body size, body shape, and age of patients. It could “learn” such knowledge by itself and handle the input images with huge differences successfully.
2. MATERIALS AND METHODS 2.1. Data acquisition A total of 278 patients with locally advanced rectal cancer receiving neoadjuvant chemoradiotherapy or radiotherapy during January 2014 to December 2016 in our department were included in this study. Patients were instructed to empty their bladder and bowels 1 h
before CT simulation. Iohexol injection (20 ml, 6 g, p.o.) added to 1 L of water was given 1 h before imaging to fill the bladder and better visualize the bowel for delineation. All patients were treated on a commercial “bellyboard” and immobilized using a thermoplastic mask. The general setup was to have the patient prone on the bellyboard. Volumetric CT data were acquired on a Somatom Definition AS 40 (Siemens Healthcare, Forchheim, Germany) or Brilliance CT Big Bore (Philips Healthcare, Best, the Netherlands) system set on helical scan
mode with or without contrast. CT images were reconstructed using a matrix size of 512×512 and thickness of 5 mm. In total, there are 25,020 CT images. Radiation oncologists contoured the CTV and OARs on the planning CT. The superior border of the CTV was the inferior margin of L5. Inferiorly, the CTV should extend to ≥3 cm below the gross disease and the ischiorectal fossa should be included for low-rectum disease. The CTV should cover the entire rectum, mesorectum, presacral space, external iliac nodes above S3, internal iliac nodes and obturator nodes. The relevant OARs were the bladder, left femoral head, right femoral head, intestine, and colon.
A total of 218 patients chosen randomly were used for training and the remaining 60 cases used for testing. The training set was used to adjust the parameters of DDCNNs, and the test set was used to assess the performance of the model. “Standard ground truth” (GT)
This article is protected by copyright. All rights reserved.
segmentations were defined as the reference segmentations generated by the radiation
Accepted Article
oncologists (who were very experienced). All the voxels that belonged to the GT segmentations of the CTV or OARs were extracted and labeled.
2.2. DDCNN model for segmentation CNNs8, 22 have become the dominant deep learning methods for medical-image analyses. We introduced a novel 2D DDCNN model to segment the CTV and OARs for radiotherapy. It was an end-to-end segmentation framework that could predict pixel-wise class labels in CT images. Figure 1 is a flowchart of the proposed model. Once the model was trained, it could segment the ROIs in 2D CT images slice by slice.
Figure 1. Flowchart of the proposed method.
This article is protected by copyright. All rights reserved.
2.2.1. Pre-processing of images
Accepted Article
For deep learning-based segmentation tasks, the final output is closely related to the quality of the input image, especially for the architecture of CNNs. However, it is difficult to distinguish tissues/organs with similar gray values, shapes, and textures in CT images.
To overcome this challenge, we proposed to pre-process the input data using an image-enhancement method to improve the image contrast. Histogram equalization (HE) is a basic method used for image enhancement, and is improved by the adaptive histogram equalization (AHE) algorithm. The AHE improves local contrast and enhances the definitions of edges in each region of an image, but tends to over-amplify noise in relatively homogeneous regions. To conquer the drawbacks of HE and AHE algorithms, we utilized the CLAHE algorithm to prevent over-amplification of noise. Image enhancement using HE, AHE and CLAHE algorithms are shown in Figure 2. With the CLAHE method, the quality of original CT images increased and the shape and texture of organs, and boundary information was highly improved.
Figure 2. Pre-processing using HE, AHE, and CLAHE algorithms.
This article is protected by copyright. All rights reserved.
Sum
Accepted Article
2.2.2. Architecture of DDCNN
D2
D1
Figure 3. Proposed DDCNN architecture.
The proposed novel DDCNN model was based on the popular VGG-16 model8, which was
designed for image classification. The network architecture is illustrated in Figure 3. Different from VGG-16, we adapted the original architecture to our segmentation task and replaced the fully connected layers with fully convolutional layers. With these deep-supervised networks and fully convolutional layers, pixel-wise segmentation in CT images could be achieved.
Moreover, organs differ from person to person based on age, body shape, and sex. To solve this problem with regard to segmentation tasks for the CTV and OARs, we used a dilated convolution filter to learn multiple-scale and multiple-level features. Similar to a convolution filter, a dilated convolution filter augments the size of the sampling space, which extracts features in larger receptive fields. Therefore, the proposed DDCNN deployed two dilated convolution modules in the architecture: the front end (D1) and back end (D2) (Figure 3). At
the front end, the dilated convolution module D1 can capture various low-level contextual features that contain the original information on intensity, texture and contour from multiple scales. At the back end, another dilated convolution module, D2, extracts high-level This article is protected by copyright. All rights reserved.
contextual features that capture holistically-aware information in larger receptive fields from
Accepted Article
feature maps. Each dilated module exploits multiple-scale features by employing multiple parallel filters with different rates. By integrating these multiple-scale features with multiple-level features, DDCNN can preserve robust information on boundaries, textures, and shapes and improve the segmentation accuracy considerably.
Table 1. The detailed architecture used in this study. layers name
type
stride
padding
dilation
output
1
1
353 × 353 × 64
2
2
353 × 353 × 128
3×3
4
4
353 × 353 × 256
3×3
8
8
353 × 353 × 512
3×3 3×3
Dilation_conv
1
3×3
1
1
3×3
2
1
Maxpool1
353 × 353 × 64 177 × 177 × 64 none
3×3
2
1
89 × 89 × 128
3×3
2
1
45 × 45 × 256
Conv1 (x2)
3×3
1
1
none
353 × 353 × 64
Maxpool2
3×3
2
1
none
177 × 177 × 64
Conv2 (x2)
3×3
1
1
none
177 × 177 × 128
Maxpool3
3×3
2
1
none
89 × 89 × 128
Conv3 (x3)
3×3
1
1
none
89 × 89 × 256
Maxpool4
3×3
2
1
none
45 × 45 × 256
Conv4 (x3)
3×3
1
1
none
45 × 45 × 512
Maxpool5
3×3
1
1
none
45 × 45 × 512
This article is protected by copyright. All rights reserved.
Accepted Article
Conv5 (x3)
3×3
1
1
none
45 × 45 × 512
Maxpool6
3×3
1
1
none
45 × 45 × 512
6
6
45 × 45 × 1024
12
12
45 × 45 × 1024
1×1
18
18
45 × 45 × 1024
1×1
24
24
45 × 45 × 1024
1×1 1×1 FC6
1
1×1
45 × 45 × 1024
1×1
45 × 45 × 1024
FC7
1
0
none
1×1
45 × 45 × 1024
1×1
45 × 45 × 1024
1×1
45 × 45 × 3
1×1
45 × 45 × 3
FC8
1
0
none
1×1
45 × 45 × 3
1×1
45 × 45 × 3
Sum Bilinear
interpolation Output
45 × 45 × 3
353 × 353 × 3
353 × 353 × 1
Specifically, Table 1 shows the architectures of the DDCNN. We deployed a novel convolutional network with 17 layers. It comprised 1 dilated convolutional layer, 16 convolutional layers, 1 bilinear interpolation, and a series of pooling options. The dilated convolutional layer included different dilation convolutions with four dilated factors of 1, 2, 4, and 8, and so had receptive fields of size 3×3, 7×7, 15×15, and 33×33 pixels. It generated 64, 128, 256, and 512 feature maps of size 353×353 through filters of size 3×3, with a stride size of 1 pixel, and a padding size of 1, 2, 4, and 8 pixels, respectively. To compensate This article is protected by copyright. All rights reserved.
low-level context information to a high-level convolutional layer, the dilated convolution
Accepted Article
branches were used to connect to other layers. However, the size of those multi-scale feature maps had to match with the higher merged layer. Therefore, the size of the dilated convolution had to be adjusted by pooling options correspondingly. Thus, the dilated convolutional features could be merged into other layers.
In addition, all the weight filters of the convolutional layers had a window size of 3×3, a stride of 1, and a padding of 1 pixel. Conv1 included two convolutional layers both generating 64 feature maps of size 353×353. Conv2 also included two convolutional layers both generating 128 feature maps of size 177×177. Conv3 included three convolutional layers all generating 256 feature maps of size 89×89. Both Conv4 and Conv5 included three convolutional layers all generating 512 feature maps of size 45×45. Pooling options could reduce the number of parameters and computation in the network, and hence control overfitting. Therefore, the layer of conv1, conv2, and conv3 were followed by a max-pooling operation with a stride of 2, a padding of 1, and window size of 3×3 for downsampling, respectively. Conv4 and conv5 were followed by a max-pooling operation with a stride of 1, a padding of 1, and window size of 3×3 for downsampling. In the layers of FC6, FC7, and FC8, the fully connected layers were replaced with convolutional layers with filters of 1×1. FC6 deployed different dilation convolutions with four dilated factors of 6, 12, 18, and 24 to extract high-level context information. FC7 used a dropout operation to select features and to avoid overfitting. The final 1×1 convolution options in FC8 carried on the pixel-level classification. Then, a sum operation generated the pixel-wise prediction. Because of the max-pooling operations, the size of the input was reduced to one-eighth of the original size in the FC8 layer. Therefore, a bilinear interpolation was used to restore the prediction map to the size of the original image. Also, the rectified linear unit (ReLU) was used as the nonlinear activation function in each convolutional layer, which can be represented as:
f ( x) max(0, x)
This article is protected by copyright. All rights reserved.
(1)
2.2.3. Training DDCNN architectures
Accepted Article
The proposed segmentation DDCNN model was implemented with Caffe (a publicly available deep learning framework). The entire process of our DDCNN comprised two processes: training and testing. The dataset is composed of data from 278 patients. We randomly picked data of 218 patients as the training dataset and used the remaining 60 cases to test the network. During the training phase, data of the 218 training patients were used to
train the network. In detail, the original 2D CT images were the inputs and the corresponding segmentation probability maps about the CTV or OARs were the outputs. To obtain a better model, we used some general methods for data enhancement, such as flip, and cut. To accelerate training, the DDCNN model was trained with the initialized parameters from the VGG-16 model trained on ImageNet and was then fine-tuned using the data of the rectal cancer. In addition, the DDCNN model was trained using backpropagation with the stochastic gradient descent (SGD) implementation of Caffe. The initial learning rate was set to 0.01 and multiplied by 0.9 after each epoch. Each model was trained for 20,000 epochs. The weight decay was set to be 0.001 and the momentum parameter was set to 0.9. The variants were trained until the training loss converged. Before each epoch, the training set was shuffled and each mini-batch (6 images) was then picked to ensure that each image was used only once in an epoch. The training iterations were 50K. The model with the highest performance was selected. The cross-entropy loss was used as the objective function for training the network. The loss was summed up over all the pixels in a mini-batch. If there was a large variation in the number of pixels in each class in the training set, then a median frequency balancing method was used to weight the loss differently based on the true class.
All computations were undertaken on a personal computer with an Intel® Core i7 processor (3.4 GHz) and a Titan Z graphics card.
This article is protected by copyright. All rights reserved.
2.3. Performance measurement
Accepted Article
When the training process was finished, the performance of the model was assessed with the
60 testing cases only. During the testing phase, all the 2D CT slices of the 60 testing cases were tested one by one. The input was the 2D CT image and the final output was pixel-level classification, which was the most likely classification label.
Performance of the proposed method was tested and compared with the segmentation of the
CTV and OARs. The Dice similarity coefficient (DSC) was used to quantify the results, and the mean and standard deviation were also calculated. It is defined as shown in Eq. (2).
DSC ( A, B)
2A B AB
(2)
where A represents the GT segmentations, B denotes the auto-segmented structure and
A B is the intersection of A and B. The DSC results in values between 0 and 1, where 0
represents no intersection at all and 1 reflects perfect overlap of structures A and B.
The performance of DDCNN was also compared with U-Net, which is used widely as the
baseline for segmentation tasks in medical images.
3. RESULTS 3.1. Performance of DDCNN and comparisons with U-Net A modified CNN called U-Net which involves connections between the convolution and deconvolutional layers has achieved great performances on medical images. Also, using
This article is protected by copyright. All rights reserved.
U-Net to evaluate segmentation results becomes a universal baseline in medical-image
Accepted Article
segmentation. Therefore, we compared the DDCNN with U-Net under an identical experimental situation. In detail, we deployed the original U-Net architecture and initial the
filters using the pre-trained Caffe model available. To carry on the experiments fairly in the training stage, U-Net had the same training configuration as that of DDCNN.
The results of this comparison are summarized in Figure 4 and Table 2. The proposed DDCNN method outperformed the U-Net on the CTV and OARs. The average DSC value of DDCNN was 3.8% higher than that of U-Net. In particular, automatic segmentation with DDCNN produced a good result for the CTV, and the value of DSC (87.7%) exceeded that of U-Net (81.9%). Among all OARs, the best values were achieved for bladder segmentation (DSC = 93.4% for DDCNN, 91.2% for U-Net). This is mainly because the bladder has good low contrast visibility and a relatively regular shape. Segmentation of the left and right femoral head also showed good agreement with the reference (DSC left: 92.1% for DDCNN, 89.6% for U-Net; DSC right: 0.923 for DDCNN, 89.2% for U-Net). The quality of the automatically generated intestine and colon was inferior to that of other structures mostly because of their complicated shapes, with mean DSC values of 65.3% and 61.8%, respectively, but remained better than that for U-Net (DSC=57.6% and 60.5%, respectively).
Figure 5 shows the visualization results in different sizes and cases. Compared with U-Net, the auto-segmented contours with DDCNN were in better agreement with the reference contours.
Thus, based on the results stated above, it seemed fair to conclude that DDCNN was a better
segmentation network architecture than the U-Net method for this task.
This article is protected by copyright. All rights reserved.
Accepted Article
Figure 4. Boxplots obtained for DSC analyses. A: DDCNN, B: U-Net
Table 2. DSC results for CTV and all OARs Left femoral
Right
head
femoral head
93.4%
92.1%
91.2%
89.6%
ROI
CTV
Bladder
DDCNN
87.7%
U-net
81.9%
Intestine
Colon
92.3%
65.3%
61.8%
89.2%
57.6%
60.5%
This article is protected by copyright. All rights reserved.
Accepted Article
Figure 5. Segmentation results for test cases shown in transverse CT slices. GT: reference segmentations; Predict: DDCNN-based segmentations; U-net: U-Net-based segmentations
3.2. Time cost The time for auto-segmentation of all the structures with the DDCNN was ≈45 s per patient
using a personal computer with an Intel® Core i7 processor (3.4G Hz) and Titan Z graphics
card.
This article is protected by copyright. All rights reserved.
4. DISCUSSION
Accepted Article
CNNs have many applications in the auto-segmentation of medical images, but delineation of the CTV and OARs in radiotherapy is manual or ABAS-based, which is the most advanced method available for medical images, but is very time-consuming.
The major benefit of
CNNs is that the training is completely end-to-end, so only the original image data are required. We compared the proposed DDCNN method with the popular U-Net (this was the first deep learning approach to CTV and OARs in radiotherapy so, for fairness, we compared it with U-Net, which is popular in medical-image analyses). The results showed that our method obtained much better DSC values than U-Net, and we can proffer an explanation. Compared to DDCNN, U-Net deploys deconvolution layers and connects the encoding layer to the decoding layer without compensating the original texture information. Therefore, U-Net is sensitive to obvious boundary regions but may lose some context information, resulting in the segmented regions with U-Net being less than those with DDCNN. DDCNN extracts multi-scale contextual features from original input image directly, which contains abundant information on context (e.g., boundaries, textures, and structure information). Therefore, the DDCNN can segment the CTV and OARs more accurately. Therefore, the DDCNN has stronger abilities in this task than U-Net.
The proposed DDCNN deployed two dilated modules to achieve multiple-level and multiple-scale feature learning. Experimental analyses demonstrated that the proposed
front-end and back-end dilated modules improved segmentation boundaries efficiently, and that structured relationships were enhanced considerably. Quantitative evaluation revealed that the segmentation accuracy was increased significantly compared with a non-dilated
architecture. Mean DSC values of all ROIs increased upon deployment of front-end and back-end dilated modules (Figure 6).
This article is protected by copyright. All rights reserved.
Accepted Article
Figure 6. Bar chart of DSC accuracy for multiple organs.
An important advantage of our approach over methods described in the literature is that it is suitable for all patients of different body size and body shape. In the experimental design, we hoped that the DDCNN could “learn” such knowledge by itself. Therefore, we chose the
training and testing sets randomly and did not separate the patients into different groups according to the size or shape of their body. Throughout the experiments, we noted good and comparable segmentation results for all patients (Figure 7). The input images displayed huge differences, but the segmentation results were pretty good. By addition of dilated convolutional modules, the proposed DDCNN architecture could be used to extract contextual features in multiple levels. Thus, the segmentation results showed that the refined boundaries and overall accuracy could be improved. Due to the lack of a common database, comparison with other methods is difficult. However, the DSC is the evaluation index used most commonly in the literature. We compared the DSC value obtained with that obtained in other studies, and found it to be at least at the same level or higher. Results for the CTV for segmentation of rectal cancer published recently have been 85.0%60, 70.0%61, 76.0%70 and 75.0%71 The DSC value of the proposed algorithm was 87.7%, which is the best result among
the results shown above. For segmentation of the bladder, DSC values have been reported to be 90.0%63, 90.0%72, 81.0%73 67.0%74, 78.3%75, and 87.0%76, whereas our proposed method demonstrated a better DSC value of 93.4%. The best segmentation of left and right femoral heads reported in the literature has yielded DSC values of 95.0%63, 91.0%73 and 92.0%74, and
our method showed comparable DSC values of 92.3% and 92.1%, respectively.
This article is protected by copyright. All rights reserved.
Accepted Article
Figure 7. Segmentation results for patients of different body size. GT: reference segmentations; Predict: DDCNN-based segmentations.
The segmentation accuracy for the intestine and colon was low, and the DSC values were 65.3% and 61.8%, respectively (Figure 8). The reasons may be as follows. Compared with other ROIs, the intestine and colon show considerable changes in shape, volume, intensity, boundary and location between patients. Contrast agents were used in all patients, but the distribution of contrast agents in the intestine differed and gas in the intestine and colon was also noted. These factors will hinder use of the DDCNN method for accurate identification of the boundaries of organs. However, studies on segmentation of the intestine or colon in radiotherapy are lacking.
This article is protected by copyright. All rights reserved.
Accepted Article
Figure 8. Segmentations of the intestine and colon. GT: reference segmentations; Predict: DDCNN-based segmentations.
Some investigations are needed to further improve the accuracy of segmentation. Due to the large model capacity, the accuracy is also expected to increase with more training data. Conversely, object segmentation in the presence of clutter and occlusions is challenging. Unfortunately, without utilizing any high-level prior information about expected objects, purely low-level information such as intensity, texture homogeneity, and strong edge contrast are not sufficient to provide the desired segmentations. In numerous studies77–81, use of prior knowledge about the shapes of the objects can improve the final reliability and accuracy of
This article is protected by copyright. All rights reserved.
the segmentation result significantly. The segmentation accuracy of a combination of deep
Accepted Article
learning methods and shape information merits further investigation.
This work was not pioneering with regard to use of a multiple-scale CNN to segment organs. Nevertheless, this work presents some novelties and may be of interest for the medical-physics community (especially for medical physicists using radiotherapy). First, we proposed a DDCNN method that auto-segments the complex CTV for radiotherapy. Second, this work was the first attempt to apply CNNs for auto-segmentation of the CTV and OARs in radiotherapy planning for rectal cancer. Third, we compared our method with U-net (which is used widely as a baseline for segmentation in medical images) and the results showed that the proposed DDCNN method outperformed the U-Net.
5. CONCLUSIONS Accurate and consistent delineation of tumor volume and OARs is particularly important in radiotherapy. However, it is labor-intensive and time-consuming. This study shows, for the first time, a method using DDCNN architecture to segment these structures in the planning
CT. Performance of DDCNN was evaluated for patients with rectal cancer and was compared with that of the popular U-Net. Quantitative results showed that DDCNN was better than U-Net and could be used to segment the CTV, bladder, and femoral heads with quite good accuracy. Moreover, it was invariant to body size, body shape, and age of the patients. Segmentation of the intestine and colon was not as good as other organs. The speed of segmentation was very fast. These data show that DDCNN can be used to improve consistency in contouring and streamlining radiotherapy workflows.
This article is protected by copyright. All rights reserved.
ACKNOWLEDGMENTS
Accepted Article
This work was supported by the National Natural Science Foundation of China [grant numbers 11605291, 11275270] and the National Key Projects of Research and Development of China [grant number 2016YFC0904600]. The authors sincerely thank Dr. Junge Zhang, Dr. Peipei Yang, Dr. Kangwei Liu and Mr. Rongliang Cheng of Institute of Automation, Chinese Academy of Sciences for data mining and editing the manuscript. We also thank the radiation oncologists in our department for delineation.
CONFLICTS OF INTEREST The authors report no conflicts of interest with this study.
References
1
L.J. Forrest, "Computed Tomography Imaging in Oncology," Veterinary Clinics of North America Small Animal Practice 46,
499-513 (2016). 2
X. Geets, J.F. Daisne, S. Arcangeli, E. Coche, P.M. De, T. Duprez, G. Nardella, V. Grégoire, "Inter-observer variability in the
delineation of pharyngo-laryngeal tumor, parotid glands and cervical spinal cord: comparison between CT-scan and MRI," Radiotherapy & Oncology 77, 25-31 (2005).
3
L. Kepka, K. Bujko, D. Garmol, J. Palucki, A. Zolciak-Siwinska, Z. Guzel-Szczepiorkowska, L. Pietrzak, K. Komosinska, A.
Sprawka, A. Garbaczewska, "Delineation variation of lymph node stations for treatment planning in lung cancer radiotherapy," Radiotherapy & Oncology 85, 450-455 (2007).
4
N.K. Jensen, D. Mulder, M. Lock, B. Fisher, R. Zener, B. Beech, R. Kozak, J. Chen, T.Y. Lee, E. Wong, "Dynamic contrast
enhanced CT aiding gross tumor volume delineation of liver tumors: an interobserver variability study," Radiotherapy & Oncology 111, 153-157 (2014). 5
P. Steenbergen, K. Haustermans, E. Lerut, R. Oyen, W.L. De, d.B.L. Van, L.G. Kerkmeijer, F.A. Pameijer, W.B. Veldhuis,
V.D.V.V.Z. Jr, "Prostate tumor delineation using multiparametric magnetic resonance imaging: Inter-observer variability and pathology validation," Radiotherapy & Oncology Journal of the European Society for Therapeutic Radiology & Oncology 115, 186-190 (2015).
This article is protected by copyright. All rights reserved.
6
L.K. Lee, S.C. Liew, J.T. Weng, A Review of Image Segmentation Methodologies in Medical Image. Lecture Notes in
Accepted Article
Electrical Engineering 315,1069-1080 (2015). 7
A. Krizhevsky, I. Sutskever, G.E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in
Neural Information Processing Systems 25,1106–1114 (2012).
8
K. Simonyan, A. Zisserman, "Very Deep Convolutional Networks for Large-Scale Image Recognition," Computer Science
(2014). 9
P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, Y. Lecun, "OverFeat: Integrated Recognition, Localization and
Detection using Convolutional Networks," Eprint Arxiv (2013). 10
R. Girshick, J. Donahue, T. Darrell, J. Malik, "Rich Feature Hierarchies for Accurate Object Detection and Semantic
Segmentation," IEEE Conference on Computer Vision and Pattern Recognition, 580-587 (2013). 11
S. Ren, K. He, R. Girshick, J. Sun, "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,"
IEEE Transactions on Pattern Analysis & Machine Intelligence, 1-1 (2016). 12
J. Long, E. Shelhamer, T. Darrell, “Fully convolutional networks for semantic segmentation,” IEEE Conference on
Computer Vision and Pattern Recognition,3431–3440 (2015).
13
L.C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A.L. Yuille, "Semantic Image Segmentation with Deep Convolutional
Nets and Fully Connected CRFs," Computer Science, 357-361 (2016). 14
S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet, "Conditional Random Fields as Recurrent Neural Networks,"
IEEE International Conference Computer Vision, 1529-1537 (2015). 15
P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, P.A. Manzagol, "Stacked Denoising Autoencoders: Learning Useful
Representations in a Deep Network with a Local Denoising Criterion," Journal of Machine Learning Research 11(12), 3371-3408 (2010). 16
G.E. Hinton , S. Osindero, Y.W. Teh, "A fast learning algorithm for deep belief nets," Neural Comput 18, 1527-1554 (2006).
17
G. Hinton, "A practical guide to training restricted Boltzmann machines," Momentum 9 (1), 926 (2010).
18
Y. Bengio, P. Simard, P. Frasconi, "Learning long-term dependencies with gradient descent is difficult," IEEE Trans Neural
Netw 5, 157-166 (1994). 19
L.C. Yann, B. Léon, B. Yoshua, H. Patrick, "Gradient-based learning applied to document recognition," Proceedings of the
IEEE 86, 2278–2324 (1998).
20
M. D. Zeiler, R. Fergus, "Visualizing and Understanding Convolutional Networks," European Conference on Computer
Vision 8689, 818-833 (2014).
21
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D.Anguelov, "Going deeper with convolutions," Computer Vision &
This article is protected by copyright. All rights reserved.
Pattern Recognition, 1-9 (2015). K. He, X. Zhang, S. Ren, J. Sun, "Deep Residual Learning for Image Recognition," Computer Science Computer Vision and
Accepted Article
22
Pattern Recognition, 770-778 (2015). 23
S. Sarraf, G. Tofighi, "Classification of Alzheimer's Disease using fMRI Data and Deep Learning Convolutional Neural
Networks," arXiv (2016). 24
C. Wang, A. Elazab, J. Wu, Q. Hu, "Lung nodule classification using deep feature fusion in chest radiography,"
Computerized Medical Imaging & Graphics 57, 10-18 (2016).
25
A. Esteva, B. Kuprel, R.A. Novoa, J. Ko, S.M. Swetter, H.M. Blau, S. Thrun, "Dermatologist-level classification of skin
cancer with deep neural networks," Nature 542, 115-118 (2017). 26
D. Yang, S. Zhang, Z. Yan, C. Tan, K. Li, D. Metaxas, "Automated anatomical landmark detection ondistal femur surface
using convolutional neural network," IEEE International Symposium on Biomedical Imaging, 17-21 (2015). 27
A. Kumar, P. Sridar, A. Quinton, R.K. Kumar, D. Feng, R. Nanan, "Plane identification in fetal ultrasound images using
saliency maps and convolutional neural networks," IEEE International Symposium on Biomedical Imaging, 791-794 (2016). 28
A. Teramoto, H. Fujita, O. Yamamuro, T. Tamaki, "Automated detection of pulmonary nodules in PET/CT images:
Ensemble false‐positive reduction using a convolutional neural network technique," Medical Physics 43(6), 2821-2827
(2016). 29
O. Ronneberger, P. Fischer, T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," International
Conference on Medical Image Computing and Computer-Assisted Intervention, 234-241 (2015). 30
F. Milletari, N. Navab, S.A. Ahmadi, "V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image
Segmentation," Fourth International Conference on 3d Vision. IEEE Computer Society, 565-571 (2016). 31
Y. Song, E.L. Tan, X. Jiang, J.Z. Cheng, D. Ni, S. Chen, B. Lei, T. Wang, "Accurate cervical cell segmentation from
overlapping clumps in pap smear images," IEEE Trans Med Imaging 36, 288-300 (2017).
32
F. Xing, Y. Xie, L. Yang, "An automatic learning-based framework for robust nucleus segmentation," IEEE Trans Med
Imaging 35 (2), 550-566 (2016). 33
H. Fu, Y. Xu, D.W.K. Wong, J.Liu, "Retinal vessel segmentation via deep learning network and fully-connected conditional
random fields," IEEE International Symposium on Biomedical Imaging, 698-701 (2016).
34
K.K. Maninis, J. Pont-Tuset, P. Arbelaez, L. Gool, "Deep retinal image understanding," Med Image Comput Comput Assist
Interv 9901 ,140–148 (2016). 35
M. Drozdzal, E. Vorontsov, G. Chartrand, S. Kadoury, C. Pal, "The importance of skip connections in biomedical image
segmentation," DLMIA 10008,179–187 (2016).
This article is protected by copyright. All rights reserved.
36
M. Shakeri, S. Tsogkas, E. Ferrante, S. Lippe, S. Kadoury, N. Paragios, I. Kokkinos, "Sub-cortical brain structure
Accepted Article
segmentation using F-CNNs," IEEE Int Symp Biomedical Imaging, 269–272 (2016). 37
K. Kamnitsas, C. Ledig, V.F. Newcombe, J.P. Simpson, A.D. Kane, D.K. Menon, "Efficient multi-scale 3D CNN with fully
connected CRF for accurate brain lesion segmentation," Medical Image Analysis 36, 61 (2017). 38
A. Birenbaum, H.Greenspan, "Longitudinal multiple sclerosis lesion segmentation using multi-view convolutional neural
networks,” International Workshop on Large-Scale Annotation of Biomedical Data and Expert Label Synthesis, 58-67 (2016). 39
T. Guo, G. Wu, L. A. Commander, S. Szary, V. Jewells, W. Lin,D. Shen, “Segmenting hippocampus from infant brains
by sparse patch matching with deep-learned features,” International Conference on Medical Image Computing and Computer-Assisted Intervention 17, 308-315 (2014). 40
H. Choi, K. H. Jin, “Fast and robust segmentation of the striatum using deep convolutional neural networks,” Journal of
neuroscience methods 274, 146-153 (2016).
41
P.V. Tran, "A fully convolutional neural network for cardiac segmentation in short-axis MRI," arXiv, 1604.00494 (2016).
42
L. Yu, Y. Guo, Y. Wang, J. Yu J, P. Chen, “Segmentation of Fetal Left Ventricle in Echocardiographic Sequences Based on
Dynamic Convolutional Neural Networks,” IEEE Transactions on Biomedical Engineering 64(8),1886-1895(2017).
43
A. Ben-Cohen, I. Diamant, E. Klang, M. Amitai, H. Greenspan, "Fully Convolutional Network for Liver Segmentation and
Lesions Detection," International Workshop on Large-Scale Annotation of Biomedical Data and Expert Label Synthesis 10008, 77-85(2016).
44
W. Thong, S. Kadoury, N. Piche, C.J. Pal, "Convolutional networks for kidney segmentation in contrast-enhanced CT
scans," Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 1-6 (2016). 45
H.R. Roth, L. Lu, A. Farag, A. Sohn, R.M. Summers, "Spatial aggregation of holistically-nested networks for automated
pancreas segmentation," Med Image Comput Comput Assist Interv 9901, 451-459 (2016). 46
R. Cheng, H.R. Roth, L. Lu, S. Wang, B. Turkbey, W. Gandler, "Active appearance model and deep learning for more
accurate prostate segmentation on MRI," Medical Imaging: Image Processing 9784, 97842I (2016).
47
K.H. Cha, L.M. Hadjiiski, R.K. Samala, H.P. Chan, R.H. Cohan, E.M. Caoili, C. Paramagul, A. Alva, A.Z. Weizer, "Bladder
cancer segmentation in CT for treatment response assessment: Application of deep-learning convolution neural network-a pilot study," Tomography 2, 421-429 (2016). 48
Y. Xu, Y. Li, M. Liu, Y. Wang, M. Lai, E.I.C. Chang, "Gland instance segmentation by deep multichannel side supervision,"
arXiv, 1607.03222 (2016).
49
R. Korez, B. Likar, F. Pernus, T. Vrtovec, "Model-based segmentation of vertebral bodies from MR images with 3D CNNs,"
Med Image Comput Comput Assist Interv 9901, 433-441 (2016). 50
A. Alansary, K. Kamnitsas, A. Davidson, R. Khlebnikov, M. Rajchl, C. Malamateniou, M.Rutherford, J.V. Hajnal, B.
This article is protected by copyright. All rights reserved.
Glocker, D. Rueckert, B. Kainz, “Fast fully automatic segmentation of the human placenta from motion corrupted MRI,”
Accepted Article
International Conference on Medical Image Computing and Computer-Assisted Intervention, 589–597 (2016). 51
J. Wang, J.D. MacKenzie,R. Ramachandran, D.Z. Chen, “A deep learning approach for semantic segmentation in histology
tissue images” International Conference on Medical Image Computing and Computer-Assisted Intervention, 176–184 (2016). 52
S. Miao, Z.J. Wang, R. Liao, "A CNN regression approach for real-time 2D/3D registration," IEEE Trans Med Imaging 35
(5), 1352-1363 (2016).
53
X. Yang, R. Kwitt, M. Niethammer, "Fast Predictive Image Registration," International Workshop on Large-Scale
Annotation of Biomedical Data and Expert Label Synthesis, 48-57 (2016). 54
W. Yang, Y. Chen, Y .Liu, L. Zhong, G. Qin, Z. Lu, "Cascade of multi-scale convolutional neural networks for bone
suppression of chest radiographs in gradient domain," Medical Image Analysis 35, 421-433 (2017). 55
D. Nie, X. Cao, Y. Gao, L. Wang, D. Shen, "Estimating CT image from MRI data using 3D fully convolutional networks,"
DLMIA 10008, 170-178 (2016). 56
X. Han, "MR-based Synthetic CT Generation using a Deep Convolutional Neural Network Method," Medical Physics 44(4),
1408-1419 (2017).
57
O. Oktay, W. Bai, M. Lee, R. Guerrero, K. Kamnitsas, J. Caballero, "Multi-input Cardiac Image Super-Resolution Using
Convolutional Neural Networks," International Conference on Medical Image Computing and Computer-Assisted Intervention, 246-254 (2016). 58
C.J. Tao, J.L. Yi, N.Y. Chen, W. Ren, J. Cheng, S. Tung, L. Kong, S.J. Lin, J.J. Pan, G.S. Zhang, "Multi-subject atlas-based
auto-segmentation reduces interobserver variation and improves dosimetric parameter consistency for organs at risk in nasopharyngeal carcinoma: A multi-institution clinical study," Radiotherapy & Oncology 115, 407-411 (2015). 59
D. Ciardo, M.A. Gerardi, S. Vigorito, A. Morra, V. Dell'Acqua, F.J. Diaz, F. Cattani, P. Zaffino, R. Ricotti, M.F. Spadea,
"Atlas-based segmentation in breast cancer radiotherapy: Evaluation of specific and generic-purpose atlases," Breast (Edinburgh, Scotland) 32, 44 (2016). 60
L.C. Anders, F. Stieler, K. Siebenlist, J. Schäfer, F. Lohr, F. Wenz, "Performance of an atlas-based autosegmentation
software for delineation of target volumes for radiotherapy of breast and anorectal cancer," Radiotherapy & Oncology Journal of the European Society for Therapeutic Radiology & Oncology 102, 68 (2012). 61
M.A. Gambacorta, C. Valentini, N. Dinapoli, L. Boldrini, N. Caria, M.C. Barba, G.C. Mattiucci, D. Pasini, B. Minsky, V.
Valentini, "Clinical validation of atlas-based auto-segmentation of pelvic volumes and normal tissue in rectal tumors using auto-segmentation computed system," Acta Oncologica 52, 1676-1681 (2013). 62
D. Li, L. Liu, J. Chen, H. Li, Y. Yin, B. Ibragimov, L. Xing, “Augmenting atlas-based liver segmentation for radiotherapy
treatment planning by incorporating image features proximal to the atlas contours," Physics in Medicine & Biology 62, 272 (2017).
This article is protected by copyright. All rights reserved.
63
W.K. Wong, L.H. Leung, D.L. Kwong, "Evaluation and optimization of the parameters used in multiple-atlas-based
Accepted Article
segmentation of prostate cancers in radiation therapy," British Journal of Radiology 89, 20140732 (2016). 64
A.A. Qazi, V. Pekar, J. Kim, J. Xie, S.L. Breen, D.A. Jaffray, "Auto-segmentation of normal and target structures in head
and neck CT images: a feature-driven model-based approach," Medical Physics 38(11), 6160-6170 (2011).
65
V.V. Gomathi, S. Karthikeyan, "A Proposed Hybrid Medoid Shift with K-Means (HMSK) Segmentation Algorithm to
Detect Tumor and Organs for Effective Radiotherapy," International Conference on Mining Intelligence and Knowledge Exploration, 139-147 (2013). 66
V.V. Gomathi, S. Karthikeyan, "A new CUT method for spliced organ segmentation in computer tomography images to
provide effective radiotherapy," International Journal of Tomography & Simulation 28(3),13-22 (2015). 67
J. Dolz, S. Ken, H.A. Leroy, N. Reyns, L. Massoptier, M. Vermandel, "Supervised machine learning method to segment the
brainstem on MRI in multicenter brain tumor treatment context," Int J Comput Assist Radiol Surg. 11(1), 43-51 (2016). 68
B. Ibragimov, L. Xing, "Segmentation of organs-at-risks in head and neck CT images using convolutional neural networks,"
Medical Physics 44(2), 547-557 (2017). 69
J. Dolz, N. Reyns, N. Betrouni, D. Kharroubi, M. Quidet, L. Massoptier, "A deep learning classification scheme based on
augmented-enhanced features to segment organs at risk on the optic region in brain cancer patients," arXiv, 1703.10480 (2017). 70
C.D. Fuller, J. Nijkamp, J.C. Duppen, C.R. Rasch, T.C. Jr, S.J. Wang, P. Okunieff, J.W. Rd, D. Baseman, S. Patel,
"Prospective randomized double-blind pilot study of site-specific consensus atlas implementation for rectal cancer target volume delineation in the cooperative group setting," International journal of radiation oncology, biology, physics 79, 481-489 (2011). 71
M.A. Gambacorta, L. Boldrini, C. Valentini, N. Dinapoli, G.C. Mattiucci, G. Chiloiro, D. Pasini, S. Manfrida, N. Caria, B.D.
Minsky, "Automatic segmentation software in locally advanced rectal cancer: READY (REsearch program in Auto Delineation sYstem)-RECTAL 02: prospective study," Oncotarget 7, 42579-42584 (2016). 72
D.P. Huyskens, P. Maingon, L. Vanuytsel, V. Remouchamps, T. Roques, B. Dubray, B. Haas, P. Kunz, T. Coradi, R.
Bühlman, "A qualitative and a quantitative analysis of an auto-segmentation module for prostate cancer," Radiotherapy & Oncology Journal of the European Society for Therapeutic Radiology & Oncology 90, 337-345 (2009). 73
G. Delpon, A. Escande, T. Ruef, J. Darréon, J. Fontaine, C. Noblet, S. Supiot, T. Lacornerie, D. Pasquier, "Comparison of
Automated Atlas-Based Segmentation Software for Postoperative Prostate Cancer Radiotherapy," Frontiers in Oncology 6 (2016). 74
J. Hwee, A.V. Louie, S. Gaede, G. Bauman, D. D'Souza, T. Sexton, M. Lock, B. Ahmad, G. Rodrigues, "Technology
assessment of automated atlas based segmentation in prostate bed contouring," Radiation Oncology 6, 110 (2011). 75
D. Li, P. Zang, X. Chai, Y. Cui, R. Li, L. Xing, "Automatic multiorgan segmentation in CT images of the male pelvis using
This article is protected by copyright. All rights reserved.
region‐specific hierarchical appearance cluster models," Medical Physics 43, 5426 (2016). V.D.S. Aj, G. Schooneveldt, S. Wognum, M.S. Hoogeman, X. Chai, L.J. Stalpers, C.R. Rasch, A. Bel, "Generic method for
Accepted Article
76
automatic bladder segmentation on cone beam CT using a patient-specific bladder shape model," Medical Physics 41, 031707 (2014). 77
J. Schmid, J. Kim, N. Magnenatthalmann, "Robust statistical shape models for MRI bone segmentation in presence of small
field of view," Medical Image Analysis 15(1), 155-168 (2011). 78
B. Ibragimov, B. Likar, F. Pernus, T. Vrtovec, "Shape Representation for Efficient Landmark-Based Segmentation in 3-D,"
IEEE Transactions on Medical Imaging 33(4), 861-874 (2014). 79
X. Qin, X. Li, Y. Liu, H. Lu, P. Yan, "Adaptive shape prior constrained level sets for bladder MR image segmentation,"
IEEE Journal of Biomedical & Health Informatics 18(5), 1707 (2014).
80
H. Ravishankar, S. Thiruvenkadam, R. Venkataramani, V. Vaidya, "Joint Deep Learning of Foreground, Background and
Shape for robust contextual segmentation," ArXiv (2016). 81
Z. Shu, C. Qi, S. Xin, C. Hu, L.Wang, Y. Zhang, "Unsupervised 3D shape segmentation and co-segmentation via deep
learning," Computer Aided Geometric Design 43(C), 39-52 (2016).
This article is protected by copyright. All rights reserved.
