Anatomical Prior-Guided Routing for 3D Lung Tumor Segmentation

Uihyeon Maeng; Jonghun Kim; Hyunjin Park

cs.CVOct 29, 2026

Anatomical Prior-Guided Routing for 3D Lung Tumor Segmentation

Authors: Uihyeon Maeng, Jonghun Kim, Hyunjin Park

Organizations: Department of Artificial Intelligence, Sungkyunkwan University, Suwon, Korea · Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, Korea

Abstract

Accurate lung tumor segmentation in computed tomography (CT) is essential for treatment planning, response assessment, and quantitative imaging analysis. However, 3D lung tumor segmentation remains challenging because tumors occupy only a small portion of the whole volume and exhibit heterogeneous appearances across anatomical locations. In patch-based training and sliding-window inference for 3D medical image segmentation, location-dependent heterogeneity is further amplified because cropped image patches provide limited global anatomical context, making it difficult for a model to adapt its feature processing according to the anatomical region from which each patch is extracted. To address this limitation, we propose an Anatomical Prior-guided Routing Module (APRM) for lung tumor segmentation. For each image patch, we obtain anatomical prior information of regional occupancy and deterministically select an anatomical routing gate. The selected gate then performs sparse feature routing over a shared expert adapter pool, enabling anatomy-specific feature adaptation while sharing expert capacity across regions. We evaluate APRM on two datasets with various backbone networks and demonstrate consistent improvements over baseline models. Furthermore, feature visualizations show that APRM produces different feature responses depending on patch location, indicating that the proposed routing mechanism effectively adapts feature processing according to anatomical context.

Figures & tables

Panel of Figure 1 — Figure 1: Distribution of lung cancer subtypes and anatomical tumor locations within the NSCLC [ 1 , 6 , 45 ] dataset. Tumor location is defined as the anatomical region (e.g., right upper lobe) showing the maximum spatial overlap with the tumor. Outside indicates areas around the lungs that do not overlap with any anatomical masks.

Figure 2: Illustration of lung tumor heterogeneity and anatomical context. Left: anatomical prior regions, including the five lung lobes and the central mediastinum. Right: different histological subtypes and lobar locations. The yellow bounding boxes show tumor regions and the red contours indicate tumor annotations.

Figure 3: Overview of proposed APRM. Anatomical masks are extracted using a pretrained TotalSegmentator [ 53 ] . An anatomical gate is assigned based on these anatomical masks. When the outside region is larger than $\tau_{\text{out}}$ , the patch is assigned to the outside gate. The selected gate computes top- $K$ sparse weights using shared experts.

	Dataset	NSCLC [ 1 , 6 , 45 ]
	Subtype	Large cell		SCC		Aden.		NOS		Average
Backbone	Module	DSC(%)	IoU(%)	DSC(%)	IoU(%)	DSC(%)	IoU(%)	DSC(%)	IoU(%)	DSC(%)	IoU(%)
nnU-Net [ 15 ]	Baseline	71.30	61.28	59.18	48.05	79.66	70.70	57.66	48.27	65.64	55.64
	+ MoE	72.52	61.86	62.54	51.10	77.77	68.79	57.65	48.95	66.66	56.55
	+ APRM	72.38	61.39	62.38	50.93	77.30	68.05	60.62	50.90	67.29	56.75
ResEnc [ 16 ]	Baseline	73.54	62.32	63.37	51.41	84.91	75.40	64.22	54.27	70.00	59.17

Table 1: Quantitative lung tumor segmentation results on the NSCLC [ 1 , 6 , 45 ] dataset with various backbones. Results are reported for each tumor subtype, where SCC denotes squamous cell carcinoma, Aden. denotes adenocarcinoma, and NOS denotes not otherwise specified. Average indicates the mean performance across all subjects. DSC and IoU are reported as percentages and higher values indicate better performance. Bold indicates the best performance.

Figure 4: Qualitative lung tumor segmentation results. Left: comparison with different modules using ResEnc as the backbone. Right: comparison with different modules using MedNeXt as the backbone. The top two rows show results for different NSCLC tumor subtypes, while the bottom row shows results on the MSD dataset. Each result shows a zoomed-in view of the yellow bounding box in the image.

	Dataset	MSD [ 2 ]
Backbone	Module	DSC(%)	IoU(%)
nnU-Net [ 15 ]	Baseline	54.86	43.09
	+ MoE	55.18	43.11
	+ APRM	58.41	46.32
ResEnc [ 16 ]	Baseline	66.38	53.47
	+ MoE	65.91	53.30

Table 2: Quantitative lung tumor segmentation results on the MSD dataset [ 2 ] . Bold indicates the best performance.

	Dataset	NSCLC [ 1 , 6 , 45 ]		MSD [ 2 ]
Category	Setting	DSC(%)	IoU(%)	DSC(%)	IoU(%)
Top- $K$ sparsity	$K=1$	70.54	59.93	65.06	52.36
Top- $K$ sparsity	$K=2$	70.88	60.10	66.72	54.02
Number of experts	$M=4$	70.80	60.21	65.74	53.29
Number of experts	$M=8$	70.88	60.10	66.72	54.02
	$\tau_{\mathrm{out}}=0.85$	69.83	59.06	66.56	53.91

Table 3: Ablation study on the NSCLC and MSD. ResEnc is used as the backbone.

Method	Total params	Router params	Infer. time (s)
nnU-Net	30.70M	–	$22.14\pm 4.80$
+ MoE	49.38M	0.017M	$48.57\pm 15.01$
+ APRM	49.48M	0.122M	$50.41\pm 15.46$

Table 4: Computational cost on the NSCLC test set using nnU-Net. Inference time reports segmentation processing only.

Figure 5: Visualization of anatomical patch selection for anatomical gate assignment. Left: an upper (superior) slice closer to the head. Right: a lower (inferior) slice closer to the abdomen. In each patch image, the upper-left label indicates the selected anatomical region, while the upper-right inset shows the distribution of anatomical masks.

Figure 6: Visualization of expert features according to anatomical location. Each image shows the voxels for which the corresponding expert is selected as the top-1 expert and the red contour denotes the tumor region. The aggregation map represents the final feature map obtained after sparse routing.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 796 papers

Segmentation, Detection and Explanation: A Unified Framework for CT Appearance Reasoning
May 15, 2026Yuyuan Liu, Can Peng, Yingyu Yang +3Image Segmentation Medical VLMs
Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford
Multimodal Routing and Region Refinement for Language-Guided Medical Image Segmentation
Sep 24, 2026Md Maklachur Rahman, Md Hasan Al Banna, Saraf Anjum +2Text-Guided Image Segmentation Medical Image Segmentation
Texas A&M University, College Station, TX 77843, USA · Independent Researcher
PGR-Net: Prior-Guided ROI Reasoning Network for Brain Tumor MRI Segmentation
Mar 23, 2026Jiacheng Lu, Hui Ding, Shiyu Zhang +1Image Segmentation Brain Tumor Segmentation
College of Information Engineering, Capital Normal University, Beijing, China · School of Artificial Intelligence, China University of Mining and Technology-Beijing, Beijing, China

May 15, 2026J/K move · Enter open · S save

Segmentation, Detection and Explanation: A Unified Framework for CT Appearance Reasoning

Yuyuan Liu, Can Peng, Yingyu Yang, Qianye Yang+2

Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford

Image Segmentation Medical VLMs Medical Image Segmentation Computed Tomography Medical Image Localization

Open paper AskPDF arXiv

Recent progress in deep learning has significantly advanced CT image analysis, particularly for segmentation tasks. However, these advances are largely confined to image-level pattern recognition, with most methods lacking explicit anatomical or contextual reasoning. Large vision-language models introduce linguistic context into image analysis, yet most approaches typically focus on a single task, which is insufficient for clinical workflow analysis that requires multiple fine-grained types of analysis, such as anatomy detection and segmentation. In this paper, we propose a unified autoregressive framework that integrates language-guided visual reasoning into CT interpretation. Our method introduces task-routing tokens that trigger detection and segmentation heads conditioned on the hidden states of a large vision-language model, enabling coherent generation of visual outputs (e.g., masks and bounding boxes) and textual reasonings. To progressively enhance localisation accuracy and semantic clarity, we further design a "closer-look" mechanism that allows the model to perform progressive coarse-to-fine visits to regions of interest under refined fields of view. To support model training and evaluation, we curated a new multimodal CT dataset containing pixel-wise masks, bounding boxes, spatial prompts, and structured descriptions for visual objects constructed through an AI-assisted annotation process with human verification. Experiments on public benchmarks demonstrate consistent improvements over the SoTA, achieving up to 1.0% Dice on BTCV and 1.7% Dice on MosMed+, while additionally providing appearance reasoning outputs. The code and dataset will be available.

Abstract

Medical Image Segmentation

Explore similar work

Segmentation, Detection and Explanation: A Unified Framework for CT Appearance Reasoning

Multimodal Routing and Region Refinement for Language-Guided Medical Image Segmentation

PGR-Net: Prior-Guided ROI Reasoning Network for Brain Tumor MRI Segmentation