Reduction of decoder convolutional blocks in Attention Gate U-Net++ to enhance computational efficiency in automatic HR-CTV delineation on CT-based cervical cancer brachytherapy images
Article Sidebar
Main Article Content
Abstract
Background: The contour of the high-risk clinical target volume (HR-CTV) is important in computed tomography (CT)-based cervical cancer brachytherapy to ensure adequate target tumor coverage while sparing radiation exposure to organs at risk (OARs). Although Attention U-Net++ has demonstrated improved segmentation performance, its complex decoder structure substantially increases computational cost and training time.
Objectives: This study aimed to develop a computationally efficient modified Attention U-Net++ architecture by reducing decoder depth while preserving segmentation performance for HR-CTV delineation.
Materials and methods: A retrospective dataset comprising 3,323 CT images from 56 patients with cervical cancer was analyzed. The proposed model eliminates the deepest decoder level to reduce redundant feature propagation and overall model complexity. Segmentation performance was evaluated using the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), 95th percentile Hausdorff Distance (HD95), Accuracy, Precision, and Recall. Computational efficiency was assessed by the number of trainable parameters, training time, and inference time, and was compared with the baseline Attention U-Net++ architecture and related models.
Results: The proposed Attention U-Net++ DL3 reduced the number of trainable parameters from 37.30 million to 8.86 million and shortened training time from 52.27 minutes to 33.11 minutes, while maintaining a comparable inference time (36.90 seconds). Geometric evaluation demonstrated a DSC of 0.9076 and IoU of 0.8376, marginally outperforming the baseline Attention U-Net++ (DSC 0.9057; IoU 0.8338). The HD95 value was 1.3959, which is comparable to existing architectures. Pixel-level classification performance remained robust, with an accuracy of 0.9996, precision of 0.9046, and recall of 0.9298. Overall, the proposed model achieved competitive segmentation performance with substantially reduced model complexity.
Conclusion: Reducing decoder convolutional blocks in Attention U-Net++ significantly improves computational efficiency while preserving segmentation accuracy. The proposed lightweight architecture provides an optimal balance between performance and resource utilization, supporting its practical implementation in CT-based cervical cancer brachytherapy, particularly in resource-limited clinical settings.
Article Details
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Personal views expressed by the contributors in their articles are not necessarily those of the Journal of Associated Medical Sciences, Faculty of Associated Medical Sciences, Chiang Mai University.
References
Priyadarshini S, Swain PK, Agarwal K, Jena D, Padhee S. Trends in gynecological cancer incidence, mortality, and survival among elderly women: a SEER study. Aging Med (Milton). 2024; 7(2): 179-88. doi: 10.1002/agm2.12297.
Burd EM. Human papillomavirus and cervical cancer. Clin Microbiol Rev. 2003; 16(1): 1-17. doi: 10.1128/CMR.16.1.1-17.2003.
Liu T, Wen S, Wang S, Yang Q, Wang X. Artificial intelligence in brachytherapy. J Radiat Res Appl Sci. 2024; 17(2): 100925. doi: 10.1016/j.jrras.2024.100925.
Dimopoulos JCA, Petrow P, Tanderup K, Petric P, Berger D, Kirisits C, et al. Recommendations from Gynaecological (GYN) GEC-ESTRO Working Group (IV): basic principles and parameters for MR imaging within the frame of image based adaptive cervix cancer brachytherapy. Radiother Oncol. 2012; 103(1): 113-22. doi: 10.1016/j.radonc.2011.12.024.
Mahantshetty U, Poetter R, Beriwal S, Grover S, Lavanya G, Rai B, et al. IBS-GEC ESTRO-ABS recommendations for CT based contouring in image guided adaptive brachytherapy for cervical cancer. Radiother Oncol. 2021; 160: 273-84. doi: 10.1016/j.radonc.2021.05.010.
van Vliet-Pérez SM, van Paassen R, Wauben LSGL, Straathof R, van den Berg NJ, Dankelman J, et al. Time-action and patient experience analyses of locally advanced cervical cancer brachytherapy. Brachytherapy. 2024; 23(3): 274-81. doi: 10.1016/j.brachy.2024.01.007.
Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015. Cham: Springer; 2015: pp 234-41. doi: 10.48550/arXiv.1505.04597.
Oktay O, Schlemper J, Folgoc LL, Lee MJ, Heinrich MP, Misawa K, et al. Attention U-Net: learning where to look for the pancreas. arXiv. 2018; 1804.03999. doi: 10.48550/arXiv.1804.03999.
Wang Y, Jian W, Zhu L, Cai C, Zhang B, Wang X. Attention-gated deep-learning-based automatic digitization of interstitial needles in high-dose-rate brachytherapy for cervical cancer. Adv Radiat Oncol. 2024; 9(1): 100941. doi: 10.1016/j.adro.2023.101340.
Zhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J. UNet++: a nested U-Net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Cham: Springer; 2018: pp 3-11. doi: 10.1007/978-3-030-00889-5_1.
Yu Y, Chen S, Wei H. Modified UNet with attention gate and dense skip connection for flow field information prediction with porous media. Flow Meas Instrum. 2023; 89: 102300. doi: 10.1016/j.flowmeasinst.2022.102300.
Zhong X, Ling W, Pan K, Liu C, Wu P, Zhang J, et al. A neural network architecture based on attention gate mechanism for 3D magnetotelluric forward modeling. Comput Geosci. 2025; 207: 106079. doi: 10.48550/arXiv.2503.11408.
Xie Z, Li X, Ma H, Wu S, Cui D. PUNet: a lightweight parallel U-Net architecture integrating Mamba-CNN for high-precision image segmentation. Sci Rep. 2025; 15: 38954. doi: 10.1038/s41598-025-22862-x.
Rajaramana S, Yang F, Zamzmi G, Xue Z, Antani S. Assessing the impact of image resolution on deep learning for TB lesion segmentation on frontal chest X-rays. Diagnostics (Basel). 2023; 13(4): 747. doi: 10.3390/diagnostics13040747.
Shorten C, Khoshgoftaar TM. A survey on image data augmentation for deep learning. J Big Data. 2019; 6: 60. doi: 10.1186/s40537-019-0197-0.
Li Z, Zhang H, Li Z, Ren Z. Residual-attention UNet++: a nested residual-attention U-Net for medical image segmentation. Appl Sci. 2022; 12(14): 7149. doi: 10.3390/app12147149.
Zhou Z, Siddiquee MMR, Tajbakhsh N, Liang J. UNet++: redesigning skip connections to exploit multiscale features in image segmentation. IEEE Trans Med Imaging. 2019; 39(6): 1856-67. doi: 10.1109/TMI.2019.2959609.
Al Qurri A, Almekkawy M. Improved UNet with attention for medical image segmentation. Sensors. 2023; 23(20): 8589. doi: 10.3390/s23208589.
Zhang L, Li M, Zhang P, Liu P. EfficientSegNet: lightweight semantic segmentation with multi-scale feature fusion and boundary enhancement. Sensors. 2025; 25: 5934. doi: 10.3390/s25195934.
Zhang Z, Liu Q, Wang Y. Road extraction by deep residual U-Net. IEEE Geosci Remote Sens Lett. 2018; 15(5): 749-53. doi: 10.1109/LGRS.2018.2802944.
Dankulchai P, Prasartseree T, Sittiwong W, Chansilpa Y, Apiwarodom N, Petsuksiri J, et al. Clinical outcome comparison between CT-guided versus all MRI-guided scenarios in brachytherapy for cervical cancer: a single-institute experience. Clin Oncol (R Coll Radiol). 2024; 36(11): e456-67. doi: 10.1016/j.clon.2024.08.003.
Qing K, Yue NJ, Hathout L, Ma C, Reyhan M, Zhu J, et al. The combined use of 2D scout and 3D axial CT images to accurately determine the catheter tips for high-dose-rate brachytherapy plans. J Appl Clin Med Phys. 2021; 22(3): 273-8. doi: 10.1002/acm2.13184.
