- Original research
- Open access
- Published:
Multi-modal co-learning with attention mechanism for head and neck tumor segmentation on 18FDG PET-CT
EJNMMI Physics volume 11, Article number: 67 (2024)
Abstract
Purpose
Effective radiation therapy requires accurate segmentation of head and neck cancer, one of the most common types of cancer. With the advancement of deep learning, people have come up with various methods that use positron emission tomography-computed tomography to get complementary information. However, these approaches are computationally expensive because of the separation of feature extraction and fusion functions and do not make use of the high sensitivity of PET. We propose a new deep learning-based approach to alleviate these challenges.
Methods
We proposed a tumor region attention module that fully exploits the high sensitivity of PET and designed a network that learns the correlation between the PET and CT features using squeeze-and-excitation normalization (SE Norm) without separating the feature extraction and fusion functions. In addition, we introduce multi-scale context fusion, which exploits contextual information from different scales.
Results
The HECKTOR challenge 2021 dataset was used for training and testing. The proposed model outperformed the state-of-the-art models for medical image segmentation; in particular, the dice similarity coefficient increased by 8.78% compared to U-net.
Conclusion
The proposed network segmented the complex shape of the tumor better than the state-of-the-art medical image segmentation methods, accurately distinguishing between tumor and non-tumor regions.
Introduction
Head and neck (H&N) cancer is one of the most common cancers and the eighth most common cause of cancer-related deaths [1, 2]. Radiation therapy cannot be performed successfully without accurate segmentation of head and neck tumors. Traditionally, radiation oncologists, medical physicists, and radiologists have manually delineated the target tumors. However, manual drawing is laborious, time-consuming, and subject to high intra- and inter-operator variabilities [3]. Therefore, automatic segmentation methods for H&N oncology are required in clinical practice and research [4].
Computed tomography (CT) images are widely used in clinical practice owing to their excellent spatial resolution and ability to provide detailed anatomical information. However, the CT intensity of H&N tumors is similar to that of adjacent tissues, which makes H&N tumor segmentation more difficult [5]. On the other hand, [18F] fluorodeoxyglucose positron emission tomography (PET) is a powerful imaging technique that can visualize metabolically active tissues, such as cancer cells [6]. Therefore, hybrid positron emission tomography-computed tomography (PET-CT) plays an important role in the diagnosis and management of H&N cancer [7, 8].
With the advancement of deep learning, various methods that use complementary information from PET-CT have been proposed. Xu et al. [9] proposed the use of cascaded V-nets for 3D volume segmentation in PET-CT imaging. In this method, the first V-Net generates a binary mask for the target object, and the mask is fed to the second V-Net along with the PET-CT images. Moreover, another method of combining features extracted from the PET and CT images via two parallel encoder networks has been proposed [10,11,12,13]. However, this method ignores the interaction of information between the two modalities during feature extraction. Xue et al. [6] achieved feature interaction across multi-modal channels by sharing features during downsampling. However, these approaches are computationally expensive because of the separation of the extraction and fusion functions and do not take full advantage of the high sensitivity of PET. In contrast, Fu et al. [14] exploited the high sensitivity of PET without co-learning PET and CT.
The main objective of this study was to achieve simple yet effective feature co-learning while fully exploiting the high sensitivity of PET. For the co-learning of multi-modality, we designed a network that learns the correlation between PET and CT features through squeeze-and-excitation normalization (SE Norm) [15] without separating the feature extraction and fusion functions. Moreover, we propose the use of a tumor region attention module (TRAM), which allows our method to fully exploit the high sensitivity of PET.
A major challenge in H&N tumor segmentation is that, unlike normal organs in the body, H&N tumors have a variety of shapes and scales. For example, as shown in Fig. 1, the shape of the tumor is irregular, causing poor generalization. Therefore, to achieve strong generalization capability, it is essential to train the network by extracting contextual information from different environments. To this end, we introduce multiscale context fusion (MSCF), which exploits contextual information from different scales.
Materials and methods
Materials
In this study, we used 224 FDG PET-CT scans of patients with head and neck tumors located in the oropharynx region, which were provided in the MICCAI HECKTOR 2021 challenge [5, 16]. The label is a CT resolution-based 3D binary description of the primary gross tumor volume (GTVt), 1 for the tumor mask, and 0 for the background. The dataset was collected from five centers (HGJ, HMR, CHUM, CHUS, and CHUP). The voxel size and resolution of the original images differed because they were acquired from different institutions.
To achieve a better and more consistent performance of the proposed network, the PET and CT images were preprocessed. First, we resampled the images to 1 \(\:\times\:\) 1 \(\:\times\:\) 1 mm3 to obtain the same voxel size. For resampling, trilinear interpolation was used for the PET and CT images, and nearest-neighbor interpolation was used for label images. The dataset we utilized provides bounding box information necessary for locating the tumor region. The dimensions of these bounding boxes are 144 \(\:\times\:\) 144 \(\:\times\:\) 144 mm3. Based on this information, the images were cropped to a size of 144 \(\:\times\:\) 144 \(\:\times\:\) 144 mm3. The intensities of CT images were clipped within the range of [-1024, 1024] Hounsfield units. Subsequently, these values were mapped to the range of [-1, 1]. The Z-score normalization was performed for PET image intensities.
Implementation details
We trained the network for 100 epochs using Python and Pytorch on a single GeForce RTX 3090 graphics processing unit. The same settings and hyperparameters were applied in all the experiments to compare the proposed model with the state-of-the-art models. The batch size was set as 1. The initial learning rate for the Adam optimizer was set to 0.001. The scheduler is a cosine annealing scheduler.
The data were split at 80:10:10 ratios for the training, validation, and test, respectively. Although data were generated from five centers, they were randomly shuffled and split without taking this into account. During training, random flips and rotations were applied for data augmentation.
Network models were comparatively evaluated using the Dice similarity coefficient (DSC), Hausdorff distance (HD), precision, and recall. The main matrices for comparison, DSC and HD, are defined as follows: (1) Dice similarity coefficient (DSC): This shows the extent of overlap between the prediction and the ground truth, ranging between 0 and 1. The DSC closer to 1 indicates that the prediction and ground truth are similar, which is defined as
where \(\:{V}_{P}\) denotes the prediction for the H&N tumor and \(\:{V}_{G}\) denotes the ground truth.
(2) Hausdorff distance (HD): It calculates the maximum distance between two segmentation surfaces and indicates the closeness between the predicted and ground truth boundaries, which is defined as
where \(\:d(p,g)\) is the Euclidean distance between points \(\:p\) and \(\:g\).
Overall networks
Figure 2 shows a schematic of the proposed network based on U-Net [17] as the backbone. The proposed network consists of three main components: TRAM, MSCF, and SE Norm. To extract features from the PET images, TRAM is separated from the backbone network, and the output of the TRAM is multiplied by the feature maps from the encoder of the backbone and inserted into the decoder. Both the MSCF and SE Norm are inserted into the backbone network. The MSCF, which exploits the different contextual information of high-level feature maps, is placed between the encoder and decoder. In deep learning, most people apply rectified linear unit (ReLU) activation function and normalization sequentially after the convolutional layer and this process are used as one block. The proposed model uses the SE Norm as a normalization of this block.
Overall architecture of our proposed model. PET and CT concatenated as two channels are input to the backbone network, and only PET is input to the TRAM. The numbers below each block indicate the number of channels. D represents the detailed block of the encoder block, and there are two types: the left one is used when the number of feature channels of input and output are different, and the right one is employed when the number of feature channels of input and output are the same
Feature encoder
We used two types of blocks in the encoder of the backbone U-Net architecture. One was used to make the channel sizes of the input and output different. The second is used when the channel sizes of the input and output are the same. Both are composed of a 3 \(\:\times\:\) 3 \(\:\times\:\) 3 convolution layer, ReLU activation function, SE Norm, and a residual layer. The difference is that only the residual layer in the former block contains a 1\(\:\times\:\)1\(\:\times\:\)1 convolution layer, a ReLU activation function, and SE Norm.
Tumor region attention module
The use of the TRAM subnetwork is proposed considering the higher sensitivity of PET images. From the PET image \(\:\varvec{P}\in\:{\mathbb{R}}^{1\:\times\:\:D\:\times\:\:H\:\times\:\:W}\), TRAM generates a tumor attention map \(\:\varvec{A}\in\:{\mathbb{R}}^{1\:\times\:\:D\:\times\:\:H\:\times\:\:W\:}\) to guide the tumor localization. A tumor attention map is incorporated into the segmentation backbone. The TRAM architecture is based on a modified U-Net consisting of a series of layers, normalizations, and activation functions. For each layer in the contracting path, a 3 \(\:\times\:\) 3 \(\:\times\:\) 3 convolution layer with a stride of two reduces the input size by half instead of 2 \(\:\times\:\) 2 \(\:\times\:\) 2 pooling layers, which is followed by batch normalization and ReLU. In the expanding path, the upsampling layer doubled the image dimensions by applying trilinear interpolation, and the number of channels was reduced by half through a 3 \(\:\times\:\) 3 \(\:\times\:\) 3 convolution layer. We applied dropout to concatenation in the decoder to reduce overfitting and add stochasticity to the network. The dropout rate is 0.5. We used the tanh activation function as the last layer to highlight meaningful features, such as tumors, and suppress less useful ones. The details of TRAM are shown in Fig. 2A.
Squeeze-and-excitation normalization
H&N tumors on PET images usually show high FDG uptake. However, some non-tumor regions also have high physiological FDG uptake. Therefore, the classification of tumors and non-tumors based solely on SUV values in PET images can be erroneous. Therefore, our network used SE Norm which enhances inter-channel relationships within the combined CT and PET feature maps to effectively blend anatomical information from CT images and metabolic information from PET images (Fig. 2C).
This module is based on the squeeze-and-excitation block, which considers the interdependencies between the channels. The inputs to the backbone network were two channels: concatenated CT and PET images. Subsequently, SE Norm was performed on each layer, and three processes were performed as follows. First, the input features were normalized by applying instance normalization. Second, because this process is performed at each channel, the output of instance normalization does not consider the information correlation between channels; therefore, to solve this problem, we generated a pair of parameters \(\:{\gamma\:}_{i},\:\:{\beta\:}_{i}\:\)from input features through SE blocks. Finally, these normalized values were scaled and shifted by \(\:{\gamma\:}_{i}\) and\(\:{\:\beta\:}_{i}\).
where \(\:{x}_{i}^{{\prime\:}}\:\)is the output of the instance normalization and \(\:{y}_{i}\) is the final output of the SE Norm.
Multi-scale context fusion
To strengthen the generalization ability of the network for diverse H&N tumor shapes, we designed an MSCF that fuses contextual information on various scales (Fig. 3) [18]. By inserting the MSCF between the encoder and decoder, we can address the issue of spatial information loss caused by consecutive pooling operations in the encoder. In this module, we extract four features at different scales by utilizing atrous convolution with different kernel sizes, which expands the receptive field of the kernels while reducing the computational burden caused by increasing the kernel size [19]. The receptive field sizes used for atrous convolution were 3,7,11 and 19, and the kernel sizes for each atrous convolution are shown in Fig. 3. The contextual information is then fused by combining the four features into one.
Feature decoder
To achieve accurate segmentation results with high resolution, the tumor attention generated by the TRAM and the SE Norm were also employed in the decoder path. The skip connection that delivers the features from the encoder is multiplied by the attention map to guide the tumor localization. Each block consists of a 3 \(\:\times\:\) 3 \(\:\times\:\) 3 convolution, ReLU, and SE Norm (Fig. 2, purple block). We also added three upsampling paths in the decoder to transmit low-resolution features, which were composed of 1 \(\:\times\:\) 1 \(\:\times\:\) 1 convolution to minimize the number of channels and trilinear interpolation to enlarge the feature map (Fig. 2, red block). Finally, to yield a segmentation probability map, the sigmoid function was applied.
Loss function
In deep learning, segmentation is usually considered as pixel-wise classification; therefore, the cross-entropy loss function is most commonly used for segmentation tasks. However, because the range of tumor size on PET-CT is considerably wide, this loss of function would not be ideal in this study. Therefore, we used the soft Dice loss function instead of the cross-entropy loss function. The Dice similarity coefficient is the most common metric for calculating the similarity between two images in computer vision, especially for medical images. The soft dice loss is defined as:
.
Our method is considered as a single integrated model and trained end-to-end using soft dice loss without applying additional constraints.
Results
Ablation experiments
Different combinations of strategies were compared to investigate the effect of incorporating the three proposed strategies (TRAM, SE Norm, and MSCF) to improve H&N tumor segmentation performance (Table 1). Adoption of TRAM, SE Norm, or MSCF resulted in better performance than the native U-Net. Among the combinations of the two strategies, the combination of TRAM and SE Norm exhibited the best performance. The proposed model with all three proposed strategies outperformed the other combinations. The DSC of the proposed model incorporating all three techniques was 8.78% greater than the DSC of the U-Net. In addition, the model reduced the number of false positives, as evidenced by its higher precision than U-Net.
Figure 4 shows an example of the segmentation results using a native U-net, a single improvement strategy (TRAM, SE Norm, or MSCA), and a combination of three strategies (proposed). In this case, only the proposed method distinguished the H&N tumor from other hot-uptake nontumor regions. Figure 5 shows that TRAM, SE Norm, and MSCF outperform native U-Net in tumor delineation.
Comparison with the state-of-the-art
Various U-Net-like models have been proposed for medical image segmentation, such as CE-Net [20], CBAM [21], CA-Net [22], MSAM [14], and Attention U-Net [23]. CE-Net combines U-Net with a novel context extractor. By placing a novel context extractor between the feature encoder and decoder in its architecture, CE-Net can avoid information loss due to successive pooling and convolution operations and highlight meaningful high-level features [20]. The network also learns variational context information using atrous convolution (See Fig. 6).
Attention U-Net, CBAM, and CA-Net use attention modules that focus on the correlation between features. In these methods, co-learning can be achieved by feeding PET and CT into the networks as two channels and focusing on the correlation between features. Attention U-Net has proposed an attention gate integrated into U-Net while minimizing computation overhead [23]. CBAM is a simple yet effective module. By applying this module to the convolution output in each block, CBAM highlights important characteristics and learns correlations in two dimensions: channels and spatial axes [21]. CA-Net uses multiple attention in convolutional neural networks to allow networks to learn spatial position, channel, and scale simultaneously [22]. To take full advantage of PET’s high sensitivity, MSAM trains a sub-network that focuses on notable regions of the PET images and a main network that provides segmentation results [14].
The proposed method focuses on the target region through co-learning, fully exploits the sensitivity of PET, and learns the variations in the shape and size of the tumors by capturing contextual information. As shown in Table 2, the proposed method outperformed the state-of-the-art models mentioned above.
Assessment of module effectiveness through visual analysis
Our research objective was to achieve simple yet effective feature co-learning by fully exploiting the high sensitivity of PET imaging. To this end, we employed two modules: TRAM and SE Norm. To understand how each module functions, we conducted further analyses.
Figure 7 displays the Tumor Attention Map generated by TRAM. This map effectively highlights areas of high uptake within PET images, indicating TRAM’s ability to focus attention on regions of interest. In essence, this suggests TRAM’s role in aiding the network in distinguishing more clearly between normal and abnormal tissues. Figure 8 visualizes the feature maps extracted from the encoder when SE Norm is applied versus when only instance normalization is used. A comparison of the two feature maps demonstrates that the use of SE Norm results in the preservation of more anatomical information. As depth of U-Net increases, feature maps that rely solely on instance normalization show a greater emergence of blank spaces. However, the application of SE Norm significantly mitigates this issue, showcasing its effectiveness in maintaining feature information at deeper network levels. These findings illustrate the pivotal role of SE Norm in enhancing feature co-learning within deep learning architectures.
Discussion
In multi-modal medical image processing [24, 25], it is useful to combine image features because the information included in each image modality is unique and complementary. Utilizing attention modules focusing on target structures and co-learning the two features are useful ways to effectively combine the image features. Appropriate utilization of PET uptake information is also important if PET is available for segment tumors. Taking all these points into account, we developed a novel deep learning-based strategy to improve the performance of H&N tumor segmentation in PET-CT. The proposed network can more effectively complement PET and CT by co-learning multi-modal features with SE Norm and guiding tumor localization with TRAM.
Attention U-Net [23] and CBAM [21] increase segmentation performance by applying additional attention modules after the convolution operation. When this method is applied to PET/CT, the features of each image are first extracted through convolution, and the extracted features are fused by the additional attention modules to perform co-learning. However, the separate extraction and fusion increases the computational complexity. In convolutional neural networks, the convolutional operation consists of a convolutional layer, an activation function, and a normalization layer. The normalization layers typically employ batch normalization or instance normalization. However, we have used SE Norm [15] as the normalization layer, which allows simultaneous feature extraction and co-learning of features. On the other hand, MSAM [14] executes learning that focuses on the high uptake of PET without any co-learning between PET and CT, which may misclassify some regions with high physiological PET uptake as tumors. In contrast to MSAM, TRAM utilizes the high sensitivity of PET and reduces false positives by combining the anatomical and physiological information of CT and PET. As shown in Table 1, the use of SE Norm with TRAM is more successful than without TRAM.
The objects of medical image segmentation, such as organs and tumors, are diverse. Normal organs usually have a similar shape, but tumors come in a variety of sizes, shapes, and locations. Therefore, networks for tumor segmentation require appropriate contextual information. CE-Net [20] devised a new context extractor and placed it between U-net encoder and decoder to learn about contextual information and highlight high-level meaningful features. In a similar way, we devised MSCF to be able to extract and learn contextual information at various scales using atrous convolution and added it to TRAM and SE Norm. Table 1 shows that when MSCF was employed with TRAM and SE Norm, DSC increased by 0.97%, which demonstrates the effectiveness of MSCF in improving the generalization ability.
We investigated the effects of combining three elements (TRAM, SE Norm, and MSCF). Learning with only one element performed worse than the two combinations, indicating that each element serves to compensate for the shortcomings of the other method. Among the two combinations, the TRAM and SE Norm combination showed the best performance. Similarly, combining all three boosted the performance improvement.
The outcomes of our study, compared with state-of-the-art (SOTA) methods, are presented in Table 2, which implies the following insights:
-
1)
Contextual information: The proposed model showed better results than CE-Net with 10.1% increase in DSC and 7.09 mm decreased in HD. Our model also delineated the tumor boundary better than CE-Net, as shown in Case 3 in Fig. 6.
-
2)
PET sensitivity: Compared to MSAM trained on the advantage of utilizing PET’s high sensitivity, our model performed better in all metrics, especially with a 3.98% increase in DSC. Figure 6 shows that false negatives and false positives of MSAM are higher than those of the proposed.
-
3)
Co-learning: Among the methods using attention modules focusing on the correlation of features, Attention U-Net showed the highest performance and CBAM showed the lowest. Our proposed model achieved 1.8% and 8.94% higher DSC than Attention U-Net and CBAM, respectively. Case studies 2 and 3 in Fig. 6 show that Attention U-Net yields more false positives than the proposed model.
However, other mothods have a better Hausdorff Distance (HD) than the proposed method in Tables 1 and 2. TRAM is the cause for the results. In order to segment tumors more accurately, TRAM generates a map that pays attention to high uptake in PET, and with the help of that, our proposed method performs segmentation. In Fig. 7, we can see a visualization of the tumor attention map, which usually gives attention to a larger area than the actual size of the tumor. Therefore, this is the reason HD becomes larger than other networks.
As demonstrated in this study, co-learning the features of multi-modal images armed with an attention mechanism that utilizes images with higher sensitivity is an effective strategy for accurate H&N tumor segmentation in PET-CT. To the best of our knowledge, this is the first study to utilize the complementary information of two modalities while exploiting the high sensitivity of PET.
Conclusions
We developed a method for H&N tumor segmentation in PET-CT using a deep neural network model that complements PET and CT more effectively by fully exploiting the high sensitivity of PET using SE Norm and guiding tumor localization using TRAM, and improved the generalization ability of the network model by utilizing contextual information with MSCF. The effectiveness of these three strategies in improving the H&N tumor segmentation was proven by comparing the performance of the proposed method with that of other state-of-the-art methods.
Data availability
All data generated or analyzed during this study are included in this published article.
References
O’rorke M, Ellison M, Murray L, Moran M, James J, Anderson L. Human papillomavirus related head and neck cancer survival: a systematic review and meta-analysis. Oral Oncol. 2012;48(12):1191–201.
Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. Cancer J Clin. 2021;71(3):209–49.
Ha S, Choi H, Paeng JC, Cheon GJ. Radiomics in oncological PET/CT: a methodological overview. Nuclear Med Mol Imaging. 2019;53:14–29.
Li X, Chen H, Qi X, Dou Q, Fu CW, Heng PA. H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes. IEEE Trans Med Imaging. 2018;37(12):2663–74.
Andrearczyk V, Oreiller V, Valli`eres M, Castelli J, Elhalawani H, Jreige M, et al. Automatic segmentation of head and neck tumors and nodal metastases in PET-CT scans. Medical imaging with deep learning. PMLR; 2020. pp. 33–43.
Xue Z, Li P, Zhang L, Lu X, Zhu G, Shen P, et al. Multi-modal co-learning for liver lesion segmentation on PET-CT images. IEEE Trans Med Imaging. 2021;40(12):3531–42.
Goel R, Moore W, Sumer B, Khan S, Sher D, Subramaniam RM. Clinical practice in PET/CT for the management of head and neck squamous cell cancer. Am J Roentgenol. 2017;209(2):289–303.
Ong SC, Sch¨oder H, Lee NY, Patel SG, Carlson D, Fury M, et al. Clinical utility of 18F-FDG PET/CT in assessing the neck after concurrent chemoradiotherapy for locoregional advanced head and neck cancer. J Nucl Med. 2008;49(4):532–40.
Xu L, Tetteh G, Lipkova J, Zhao Y, Li H, Christ P et al. Automated whole-body bone lesion detection for multiple myeloma on 68 Ga-Pentixafor PET/CT imaging using deep learning methods. Contrast media & molecular imaging. 2018;2018.
Zhao X, Li L, Lu W, Tan S. Tumor co-segmentation in PET/CT using multi-modality fully convolutional neural network. Phys Med Biol. 2018;64(1):015011.
Zhong Z, Kim Y, Plichta K, Allen BG, Zhou L, Buatti J, et al. Simultaneous cosegmentation of tumors in PET-CT images using deep fully convolutional networks. Med Phys. 2019;46(2):619–33.
Jin D, Guo D, Ho TY, Harrison AP, Xiao J, Tseng CK et al. Accurate esophageal gross tumor volume segmentation in PET/CT using two-stream chained 3D deep network fusion. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. Springer; 2019. pp. 182 – 91.
Kumar A, Fulham M, Feng D, Kim J. Co-learning feature fusion maps from PET-CT images of lung cancer. IEEE Trans Med Imaging. 2019;39(1):204–17.
Fu X, Bi L, Kumar A, Fulham M, Kim J. Multimodal spatial attention module for targeting multimodal PET-CT lung tumor segmentation. IEEE J Biomedical Health Inf. 2021;25(9):3507–16.
Iantsen A, Jaouen V, Visvikis D, Hatt M. Squeeze-and-excitation normalization for brain tumor segmentation. In: Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 6th International Workshop, BrainLes 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Revised Selected Papers, Part II 6. Springer; 2021. pp. 366 – 73.
Andrearczyk V, Oreiller V, Boughdad S, Rest CCL, Elhalawani H, Jreige M, et al. Overview of the HECKTOR challenge at MICCAI 2021: automatic head and neck tumor segmentation and outcome prediction in PET/CT images. 3D head and neck tumor segmentation in PET/CT challenge. Springer; 2021. pp. 1–37.
C¸ i¸cek O, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U- ¨ Net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17–21, 2016, Proceedings, Part II 19. Springer; 2016. pp. 424 – 32.
Chen LC, Papandreou G, Kokkinos I, Murphy K, Yuille AL, Deeplab. Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans Pattern Anal Mach Intell. 2017;40(4):834–48.
Chen LC, Papandreou G, Schroff F, Adam H. Rethinking atrous convolution for semantic image segmentation. arXiv Preprint arXiv:170605587. 2017.
Gu Z, Cheng J, Fu H, Zhou K, Hao H, Zhao Y, et al. Ce-net: Context encoder network for 2d medical image segmentation. IEEE Trans Med Imaging. 2019;38(10):2281–92.
Woo S, Park J, Lee JY, Kweon IS, Cbam. Convolutional block attention module. In: Proceedings of the European conference on computer vision (ECCV); 2018. pp. 3–19.
Gu R, Wang G, Song T, Huang R, Aertsen M, Deprest J, et al. CANet: comprehensive attention convolutional neural networks for explainable medical image segmentation. IEEE Trans Med Imaging. 2020;40(2):699–711.
Oktay O, Schlemper J, Folgoc LL, Lee M, Heinrich M, Misawa K et al. Attention u-net: Learning where to look for the pancreas. arXiv. 2018. arXiv preprint arXiv:180403999. 1804.
Lee MS, Hwang D, Kim JH, Lee JS. Deep-dose: a voxel dose estimation method using deep convolutional neural network for personalized internal dosimetry. Sci Rep. 2019;9(1):10308.
Park J, Bae S, Seo S, Park S, Bang JI, Han JH, et al. Measurement of glomerular filtration rate using quantitative SPECT/CT and deep-learning based kidney segmentation. Sci Rep. 2019;9(1):4223.
Acknowledgements
Not applicable.
Funding
This work was supported by the Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137953, RS-2020-KD000028) and by grant from the National Research Foundation of Korea (NRF) funded by the Korean Ministry of Science and ICT (Grant No. 2020M2D9A1093989).
Author information
Authors and Affiliations
Contributions
All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by MJC, DH, SYY, and JSL. The first draft of the manuscript was written by MJC and all authors commented on previous versions of the manuscript. JSL did supervision and funding acquisition. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
No ethical approval was required as this study utized open dataset.
Consent for publication
Not applicable.
Competing interests
The authors have no relevant financial or non-financial interests to disclose.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Cho, M.J., Hwang, D., Yie, S.Y. et al. Multi-modal co-learning with attention mechanism for head and neck tumor segmentation on 18FDG PET-CT. EJNMMI Phys 11, 67 (2024). https://doi.org/10.1186/s40658-024-00670-y
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s40658-024-00670-y