Objective Accurate three-dimensional segmentation of lung tumors from computed tomography (CT) images is important for tumor diagnosis, treatment planning, staging, and therapeutic response assessment. However, clinical CT images usually have anisotropic spatial resolution, with higher resolution in the in-plane directions than in the inter-slice direction. Sparse sampling along the z-axis may lead to discontinuous tumor structures and inaccurate boundary localization. In addition, heterogeneous intensity, irregular morphology, ambiguous boundaries, and low contrast with surrounding tissues make tumor features vulnerable to background interference, resulting in missed segmentation and false-positive predictions. Existing three-dimensional networks generally use similar feature extraction strategies along different spatial axes, while conventional skip connections provide limited interaction among non-adjacent multi-scale features. To address these problems, a three-dimensional lung tumor segmentation network based on inter-slice heterogeneous synergy and tridirectional guided fusion, termed HTG-Net, was developed.

Methods HTG-Net adopted an encoder–decoder architecture. An inter-slice heterogeneous synergy module, named IHSM, was embedded into the encoder to model the differences between in-plane and inter-slice information. IHSM combined a three-dimensional spatial continuity branch with a two-dimensional pseudo-plane interaction branch. The former extracted volumetric context and anatomical continuity, whereas the latter rearranged depth-related information and modeled local relationships among neighboring slices. The two branches were integrated through feature mixing and adaptive gated fusion, allowing the network to preserve in-plane texture while improving inter-slice representation. A tridirectional guided fusion module, named TGFM, was introduced after the encoder to promote interaction among multi-scale features. TGFM used DownFlow, UpFlow, and ForwardFlow to transmit shallow details, deep semantics, and current-level features across encoder stages. A multi-scale guided efficient channel attention module, named MSG-ECA, was further used to estimate the spatial contribution of features at different scales and recalibrate channel responses, thereby enhancing tumor-related information and suppressing background interference. The network was evaluated on a clinical lung tumor CT dataset containing 307 patients, the MSD-Lung dataset containing 63 cases, and the RIDER-LungCT-Seg dataset containing 62 CT scans from 31 patients. Five-fold cross-validation was used for evaluation, and cross-dataset testing was performed on RIDER-LungCT-Seg.

Results and Discussions On the clinical dataset, HTG-Net achieved a Dice similarity coefficient (DSC) of 83.89%, a 95th-percentile Hausdorff distance (HD95) of 4.40 mm, a sensitivity of 80.84%, and a positive predictive value of 81.93%. Compared with nnU-Net, DSC increased by 1.80 percentage points and HD95 decreased by 1.03 mm. On the MSD-Lung dataset, HTG-Net achieved a DSC of 78.87% and an HD95 of 6.29 mm, improving DSC by 0.78 percentage points over nnU-Net and obtaining the lowest HD95 among the comparison methods. On the external RIDER-LungCT-Seg dataset, HTG-Net achieved a DSC of 69.17%, an HD95 of 7.22 mm, and a sensitivity of 62.36%. Compared with nnU-Net, DSC increased by 0.42 percentage points, HD95 decreased by 1.30 mm, and sensitivity increased by 0.93 percentage points. Ablation experiments showed that IHSM improved boundary representation, whereas the tridirectional flow and MSG-ECA enhanced cross-scale feature fusion. Visual results demonstrated that HTG-Net better preserved narrow connections, irregular lobulations, sharp boundaries, and small protruding structures. The network also achieved a favorable balance between segmentation performance and computational complexity.

Conclusions HTG-Net improves lung tumor segmentation in anisotropic CT images by combining heterogeneous inter-slice modeling with guided multi-scale fusion. Experiments on three datasets demonstrated improved segmentation accuracy, boundary localization, and cross-dataset generalization. The proposed network provides a potential approach for automated lung tumor segmentation, while future work will focus on refined boundary modeling, model lightweighting, and broader clinical evaluation.