Lung tumor segmentation improves with anatomy guided feature routing
Anatomical Prior-Guided Routing for 3D Lung Tumor Segmentation
Computer Vision and Pattern Recognition
Summary
Lung tumors are hard to spot accurately in 3D CT scans because they are small and look different depending on where they are in the lung. The authors created a method that helps the computer focus differently on parts of the scan based on where the tumor is likely located. Their method uses anatomical information to guide how the computer processes features, which helps it better recognize tumors in different lung regions. Tests show this approach improves tumor detection compared to previous methods.
What this means in practice
- •For medical imaging teams: Improve lung tumor detection accuracy by using anatomy-guided feature processing in 3D CT segmentation workflows.
- •For radiology software developers: Develop software modules that adapt feature extraction based on lung anatomy to enhance tumor segmentation performance.
Authors
Uihyeon Maeng, Jonghun Kim, Hyunjin Park
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.