A multi-scale supervision and residual feedback optimization algorithm for improving optic chiasm and optic nerve segmentation accuracy in nasopharyngeal carcinoma CT images

doi:10.12122/j.issn.1673-4254.2025.03.21

Abstract

Abstract:

Objective We propose a novel deep learning segmentation algorithm (DSRF) based on multi-scale supervision and residual feedback strategy for precise segmentation of the optic chiasm and optic nerves in CT images of nasopharyngeal carcinoma (NPC) patients. Methods We collected 212 NPC CT images and their ground truth labels from SegRap2023, StructSeg2019 and HaN-Seg2023 datasets. Based on a hybrid pooling strategy, we designed a decoder (HPS) to reduce small organ feature loss during pooling in convolutional neural networks. This decoder uses adaptive and average pooling to refine high-level semantic features, which are integrated with primary semantic features to enable network learning of finer feature details. We employed multi-scale deep supervision layers to learn rich multi-scale and multi-level semantic features under deep supervision, thereby enhancing boundary identification of the optic chiasm and optic nerves. A residual feedback module that enables multiple iterations of the network was designed for contrast enhancement of the optic chiasm and optic nerves in CT images by utilizing information from fuzzy boundaries and easily confused regions to iteratively refine segmentation results under supervision. The entire segmentation framework was optimized with the loss from each iteration to enhance segmentation accuracy and boundary clarity. Ablation experiments and comparative experiments were conducted to evaluate the effectiveness of each component and the performance of the proposed model. Results The DSRF algorithm could effectively enhance feature representation of small organs to achieve accurate segmentation of the optic chiasm and optic nerves with an average DSC of 0.837 and an ASSD of 0.351. Ablation experiments further verified the contributions of each component in the DSRF method. Conclusion The proposed deep learning segmentation algorithm can effectively enhance feature representation to achieve accurate segmentation of the optic chiasm and optic nerves in CT images of NPC.

Key words: nasopharyngeal carcinoma, optic chiasm and optic nerve segmentation, hybrid pooling strategy, deep supervision, residual feedback

Jinyu LIU, Shujun LIANG, Yu ZHANG. A multi-scale supervision and residual feedback optimization algorithm for improving optic chiasm and optic nerve segmentation accuracy in nasopharyngeal carcinoma CT images[J]. Journal of Southern Medical University, 2025, 45(3): 632-642.

Figures/Tables 12

Fig.1 CT images and ground truth labels for segmentation of the optic chiasm and optic nerve in CT images from 3 datasets. The optic chiasm is shown in red, the left optic nerve in green, and the right optic nerve in blue.

Fig.2 Deep supervision residual feedback network framework. The orange components represent the encoder, the green components represent the decoder based on the Hybrid Pooling Strategy (HPS), the blue components represent the Residual Feedback Module (RFM), and the cyan components represent the Multi-scale Deep Supervision Layers (DSL). The top-right corner shows a magnified view of the residual information map and residual mask. The bottom-right corner illustrates the structure of the residual unit.

Fig.3 Decoder based on hybrid pooling strategy. A: Expanded Pyramid Pooling Module. B: Deep Pooling Layer. C: Pooling Unit.

Fig.4 Multi-scale deep supervision layers.

Fig.5 Visualized results of the ablation experiments on the internal test set for DSRF. The ground truth labels are shown in red, and the segmentation results of the model are shown in green.

Fig.6 Examples of semantic feature heatmaps in the decoder blocks. w-HPS represents the DSRF model, o-HPS represents the DSL+RFM model, and De4 to De1 correspond to decoder blocks from deep to shallow layers. GT denotes the ground truth labels. The heatmap colors range from white to red, indicating semantic feature values from low probability to high probability.

Fig.7 Two examples of residual expression. The red contour in the first row represents the edge of the ground truth label. In the second row, the residual information color changes (from blue to red) indicate low to high probabilities.

Tab.1 Average dice similarity coefficient of DSRF in ablation experiments on the internal and external test sets

HPS	DSL	RFM	Internal Test			External Test
HPS	DSL	RFM	aaChiasmaa	OpticNrv_L	OpticNrv_R	aaChiasmaa	OpticNrv_L	OpticNrv_R
			0.689	0.789	0.793	0.537	0.711	0.706
√			0.729	0.836	0.838	0.659	0.768	0.773
	√		0.706	0.808	0.801	0.594	0.730	0.736
		√	0.701	0.796	0.796	0.564	0.718	0.717
	√	√	0.712	0.818	0.821	0.601	0.747	0.751
√		√	0.743	0.852	0.858	0.672	0.788	0.789
√	√		0.752	0.858	0.857	0.688	0.801	0.813
√	√	√	0.764	0.872	0.874	0.708	0.819	0.828

Tab.2 Average symmetric surface distance of DSRF in ablation experiments on the internal and external test sets

HPS	DSL	RFM	Internal Test			External Test
HPS	DSL	RFM	aaChiasmaa	OpticNrv_L	OpticNrv_R	aaChiasmaa	OpticNrv_L	OpticNrv_R
			0.875	0.450	0.451	1.386	0.728	0.783
√			0.668	0.326	0.343	0.925	0.537	0.541
	√		0.838	0.437	0.447	1.168	0.626	0.646
		√	0.867	0.448	0.452	1.327	0.658	0.679
	√	√	0.768	0.366	0.414	1.068	0.568	0.614
√		√	0.657	0.279	0.270	0.859	0.447	0.440
√	√		0.640	0.247	0.270	0.690	0.374	0.388
√	√	√	0.601	0.220	0.232	0.615	0.334	0.333

Tab.3 Quantitative results of different iteration counts for DSRF

Method	DSC			ASSD
Method	aaChiasmaa	OpticNrv_L	OpticNrv_R	aaChiasmaa	OpticNrv_L	OpticNrv_R
DSRF⁰	0.752	0.858	0.857	0.640	0.247	0.270
DSRF¹	0.764	0.872	0.874	0.601	0.220	0.232
DSRF²	0.759	0.875	0.876	0.598	0.214	0.220

Tab.4 Segmentation results of DSRF and other methods on the internal test set

Method	DSC			ASSD
Method	aaChiasmaa	OpticNrv_L	OpticNrv_R	aaChiasmaa	OpticNrv_L	OpticNrv_R
nnU-Net^[13]	0.721	0.809	0.824	0.632	0.354	0.363
PoolNet^[17]	0.722	0.838	0.834	0.676	0.330	0.353
STU-Net^[18]	0.732	0.802	0.828	0.640	0.422	0.368
UMamba^[19]	0.733	0.816	0.814	0.769	0.371	0.409
RF-Net^[9]	0.735	0.863	0.869	0.807	0.246	0.242
DSRF	0.764	0.872	0.874	0.601	0.220	0.232

Fig.8 Visual comparison of the segmentation results between the ground truth (red) and the proposed framework (green) for two randomly selected testing subjects from the internal and external test sets.

References 32

1	Du XJ, Wang GY, Zhu XD, et al. Refining the 8th edition TNM classification for EBV related nasopharyngeal carcinoma[J]. Cancer Cell, 2024, 42(3): 464-73.e3.
2	陈美宁, 刘懿梅, 彭应林, 等. 不同级别肿瘤中心医师对鼻咽癌调强放疗靶区和危及器官勾画差异比较[J]. 中国医学物理学杂志, 2024, 41(3): 265-72.
3	Guo DZ, Jin DK, Zhu Z, et al. Organ at risk segmentation for head and neck cancer using stratified learning and neural architecture search[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020: 4223-32.
4	Guo H, Wang J, Xia X, et al. The dosimetric impact of deep learning-based auto-segmentation of organs at risk on nasopharyngeal and rectal cancer[J]. Radiat Oncol, 2021, 16(1): 113.
5	Costea M, Zlate A, Durand M, et al. Comparison of atlas-based and deep learning methods for organs at risk delineation on head-and-neck CT images using an automated treatment planning system[J]. Radiother Oncol, 2022, 177: 61-70.
6	Peng YL, Liu YM, Shen GZ, et al. Improved accuracy of auto-segmentation of organs at risk in radiotherapy planning for nasopharyngeal carcinoma based on fully convolutional neural network deep learning[J]. Oral Oncol, 2023, 136: 106261.
7	Luan S, Wei C, Ding Y, et al. PCG-net: feature adaptive deep learning for automated head and neck organs-at-risk segmentation[J]. Front Oncol, 2023, 13: 1177788.
8	Liu P, Sun Y, Zhao X, et al. Deep learning algorithm performance in contouring head and neck organs at risk: a systematic review and single-arm meta-analysis[J]. Biomed Eng Online, 2023, 22(1): 104.
9	Wang K, Liang SJ, Zhang Y. Residual feedback network for breast lesion segmentation in ultrasound image[M]//Medical Image Computing and Computer Assisted Intervention – MICCAI 2021. Cham: Springer International Publishing, 2021: 471-81.
10	Luo XD, Fu J, Zhong YX, et al. SegRap2023: a benchmark of organs-at-risk and gross tumor volume Segmentation for Radiotherapy Planning of Nasopharyngeal Carcinoma[J]. Med Image Anal, 2025, 101: 103447.
11	Automatic Structure Segmentation for Radiotherapy Planning Challenge 2019. The MICCAI 2019 Challenge[OL]. Retrieved from
12	Podobnik G, Strojan P, Peterlin P, et al. HaN-Seg: The head and neck organ-at-risk CT and MR segmentation dataset[J]. Med Phys, 2023, 50(3): 1917-27.
13	Isensee F, Jaeger PF, Kohl SAA, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation[J]. Nat Meth, 2021, 18: 203-11.
14	He KM, Zhang XY, Ren SQ, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016. Las Vegas, NV, USA. IEEE, 2016: 770-778.
15	Kingma DP, Ba J. Adam: A method for stochastic optimization [J]. arXiv preprint arXiv:, 2014.
16	Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks[J]. Commun ACM, 2017, 60(6): 84-90.
17	Liu JJ, Hou QB, Cheng MM, et al. A simple pooling-based design for real-time salient object detection[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 15-20, 2019. Long Beach, CA, USA. IEEE, 2019: 3917-3926.
18	Huang ZY, Ye J, Wang HY, et al. Evaluating STU-net for brain tumor segmentation[M]//Brain Tumor Segmentation, and Cross-Modality Domain Adaptation for Medical Image Segmentation. Cham: Springer Nature Switzerland, 2024: 140-51.
19	Ma J, Li FF, Wang B. U-mamba: enhancing long-range dependency for biomedical image segmentation[EB/OL]. 2024: 2401.04722. .
20	Wenderott K, Krups J, Zaruchas F, et al. Effects of artificial intelligence implementation on efficiency in medical imaging: a systematic literature review and meta-analysis[J]. NPJ Digit Med, 2024, 7: 265.
21	赖建军, 陈丽婷, 胡海丽, 等. 基于深度学习自动勾画在鼻咽癌调强放射治疗计划中的系统性评价研究[J]. 中国现代医药杂志, 2023, 25(10): 24-30.
22	黄新, 王新卓, 薛涛, 等. 鼻咽癌放射治疗危及器官自动勾画的几何和剂量学分析[J]. 生物医学工程与临床, 2024, 28(1): 26-34.
23	Su ZY, Siak PY, Lwin YY, et al. Epidemiology of nasopharyngeal carcinoma: current insights and future outlook[J]. Cancer Metastasis Rev, 2024, 43(3): 919-39.
24	Azad R, Aghdam EK, Rauland A, et al. Medical image segmentation review: the success of U-net[J]. IEEE Trans Pattern Anal Mach Intell, 2024, 46(12): 10076-95.
25	Liu JJ, Hou Q, Liu ZA, et al. PoolNet+: exploring the potential of pooling for salient object detection[J]. IEEE Trans Pattern Anal Mach Intell, 2023, 45(1): 887-904.
26	Hu K, Chen W, Sun Y, et al. PPNet: Pyramid pooling based network for polyp segmentation[J]. Comput Biol Med, 2023, 160: 107028.
27	Wu YH, Liu Y, Zhan X, et al. P2T: pyramid pooling transformer for scene understanding[J]. IEEE Trans Pattern Anal Mach Intell, 2023, 45(11): 12760-71.
28	Wang LW, Lee CY, Tu ZW, et al. Training deeper convolutional networks with deep supervision[EB/OL]. 2015: 1505.02496. .
29	Zhang LF, Chen X, Zhang JB, et al. Contrastive deep supervision[M]//Computer Vision-ECCV 2022. Cham: Springer Nature Switzerland, 2022: 1-19.
30	Ahmad S, Ullah Z, Gwak J. Multi-teacher cross-modal distillation with cooperative deep supervision fusion learning for unimodal segmentation[J]. Knowl Based Syst, 2024, 297: 111854.
31	Umer MJ, Sharif MI, Kim J. Breast cancer segmentation from ultrasound images using multiscale cascaded convolution with residual attention-based double decoder network[J]. IEEE Access, 2024, 12: 107888-902.
32	Wang J, Liang J, Xiao Y, et al. TaiChiNet: negative-positive cross-attention network for breast lesion segmentation in ultrasound images[J]. IEEE J Biomed Health Inform, 2024, 28(3): 1516-27.

[1]	Lu TAO, Zhuoli WEI, Yueyue WANG, Ping XIANG. CEACAM6 inhibits proliferation and migration of nasopharyngeal carcinoma cells by suppressing epithelial-mesenchymal transition [J]. Journal of Southern Medical University, 2025, 45(3): 566-576.
[2]	Xiaofan CONG, Teng CHEN, Shuo LI, Yuanyuan WANG, Longyun ZHOU, Xiaolong LI, Pei ZHANG, Xiaojin SUN, Surong ZHAO. Dihydroartemisinin enhances sensitivity of nasopharyngeal carcinoma HNE1/DDP cells to cisplatin-induced apoptosis by promoting ROS production [J]. Journal of Southern Medical University, 2024, 44(8): 1553-1560.
[3]	Yuanyuan WANG, Teng CHEN, Xiaofan CONG, Yiran LI, Rui CHEN, Pei ZHANG, Xiaojin SUN, Surong ZHAO. Pristimerin enhances cisplatin-induced apoptosis in nasopharyngeal carcinoma cells via ROS-mediated deactivation of the PI3K/AKT signaling pathway [J]. Journal of Southern Medical University, 2024, 44(5): 904-912.
[4]	Yue HU, Yu ZENG, Linjing WANG, Zhiwei LIAO, Jianming TAN, Yanhao KUANG, Pan GONG, Bin QI, Xin ZHEN. Performance of multi-modality and multi-classifier fusion models for predicting radiation-induced oral mucositis in patients with nasopharyngeal carcinoma [J]. Journal of Southern Medical University, 2024, 44(12): 2434-2442.
[5]	ZHANG Haoxuan, LU Jin, JIANG Chengyi, FANG Meifang. Construction and evaluation of an artificial intelligence-based risk prediction model for death in patients with nasopharyngeal cancer [J]. Journal of Southern Medical University, 2023, 43(2): 271-279.
[6]	HU Tong, GOU Wenfeng, REN Zhonghao, LIU Gaiting, LI Yiliang, ZUO Daiying, HOU Wenbin. Icaritin increases radiosensitivity of nasopharyngeal carcinoma cells by regulating iron death [J]. Journal of Southern Medical University, 2023, 43(10): 1665-1673.
[7]	ZHANG Hengyi, PANG Jinlong, ZHANG Yuhan, MA Yue, FAN Fangtian, LIU Hao. AZD9291 suppresses proliferation and migration of nasopharyngeal carcinoma cells by inhibiting the PI3K-AKT-mTOR pathway [J]. Journal of Southern Medical University, 2022, 42(9): 1403-1409.
[8]	ZHOU Lanzhu, WU Jun, ZHANG Mingjie, ZHAO Bao, MA Shiyin. PRMT1 inhibits apoptosis of nasopharyngeal carcinoma cells by promoting RRM2 expression [J]. Journal of Southern Medical University, 2022, 42(12): 1783-1790.
[9]	GAO Lili, ZHANG Xiong, DOU Siyu, YUE Xiaoding, YANG Jieling. Interference of long noncoding RNA FOXCUT inhibits epithelial-mesenchymal transformation and induces mitochondrial injury in nasopharyngeal carcinoma cells [J]. Journal of Southern Medical University, 2021, 41(9): 1334-1341.
[10]	. TRIM59 regulates invasion and migration of nasopharyngeal carcinoma cells by targeted modulation of PPM1B [J]. Journal of Southern Medical University, 2021, 41(7): 1030-1036.
[11]	. Effect of exosomes derived from human Epstein-Barr virus-positive nasopharyngeal carcinoma cells on lymphangiogenesis and lymph node metastasis [J]. Journal of Southern Medical University, 2020, 40(12): 1776-1783.
[12]	. Down-regulation of miR-205-5p enhances pro-apoptotic effect of 3-bromopyruvate on human nasopharyngeal carcinoma CNE2Z cells [J]. Journal of Southern Medical University, 2019, 39(10): 1166-.
[13]	. Quantitative proteomics and differential signal enrichment in nasopharyngeal carcinoma cells with or without SETD2 gene knockout [J]. Journal of Southern Medical University, 2019, 39(10): 1191-.
[14]	. YAP1 knockdown suppresses the proliferation, migration and invasion of human nasopharyngeal carcinoma cells [J]. Journal of Southern Medical University, 2019, 39(03): 286-.
[15]	. Long non-coding RNA XIST modulates cisplatin resistance by altering PDCD4 and Fas-L expressions in human nasopharyngeal carcinoma HNE1 cells in vitro [J]. Journal of Southern Medical University, 2019, 39(03): 357-.