Why accurate segmentation matters. Pulmonary nodules are small abnormal growths in the lung tissue that can be early indicators of lung cancer. Accurate segmentation, meaning precisely delineating the boundary between nodule and surrounding lung tissue in CT scans, is essential for measuring nodule volume, tracking growth over time, and guiding biopsy or treatment planning. Manual segmentation by radiologists is time-consuming and suffers from inter-observer variability.
The difficulty of complex nodule types. Not all nodules are equally easy to segment. Adherent nodules are attached to the pleural wall or blood vessels, making their boundaries indistinct from surrounding structures. Calcified nodules have unusually high CT density that can mislead intensity-based segmentation methods. These two types remain a significant source of errors in automated systems.
Why 3D matters for volumetric CT data. CT scans are inherently three-dimensional. A nodule is a 3D object that appears across multiple adjacent image slices. Methods that process each 2D slice independently miss spatial relationships between slices, leading to inconsistent boundaries and missed context. True 3D volumetric segmentation methods treat the CT as a full volume, enabling the model to consider context in all three spatial dimensions simultaneously.
The limitations of current attention mechanisms. Standard attention mechanisms in medical image segmentation, such as coordinate attention and selective kernel modules, are either computationally expensive or do not fully leverage cross-dimensional spatial relationships in 3D volumes. This gap motivates the development of more efficient grouped attention approaches that maintain accuracy while reducing parameter count.
V-Net as the foundation. V-Net is a fully 3D convolutional neural network designed specifically for volumetric medical image segmentation. It uses a symmetric encoder-decoder architecture where the encoder progressively downsamples the input to capture high-level semantic features, and the decoder upsamples back to the original resolution to produce pixel-wise segmentation maps. Skip connections link corresponding encoder and decoder levels, allowing the decoder to reuse spatial detail from earlier layers.
Input and output dimensions. In this study, the V-Net receives a single-channel 3D grayscale CT patch of size 1 by 16 by 96 by 96 voxels. It produces a 2-channel output of the same spatial size, with one channel representing the nodule class and the other representing background, enabling per-voxel binary classification.
Where standard V-Net falls short. The original V-Net lacks spatial attention mechanisms, meaning it treats all regions of the input volume equally regardless of their relevance to the segmentation target. This causes it to struggle with nodules adjacent to structures of similar density, such as vessels and pleural walls, where the absence of focused attention leads to over-segmentation or under-segmentation at boundaries.
The DGCA V-Net improvement strategy. This work augments the standard V-Net by inserting three novel attention modules at specific architectural locations: the Global Grouped Coordinate Attention (GGCA) in the encoder to enhance feature discrimination, the Grouped Split Attention (GSA) in the decoder to preserve fine boundary detail during reconstruction, and the Dual-input-guided feature Aggregation (DGA) on skip connections to selectively fuse encoder and decoder features.
Why coordinate-aware attention helps in 3D. Standard squeeze-and-excitation attention compresses the entire 3D feature map into a single global vector per channel, discarding all spatial positional information. Coordinate attention preserves positional information by computing separate attention along each spatial dimension. In 3D volumes this becomes especially important because nodule features vary substantially across height, width, and depth directions.
The grouping innovation for efficiency. The GGCA module divides the input feature channels into groups before computing attention. Within each group, cross-dimension attention is computed by pooling along each of the three spatial axes independently. The resulting dimension-specific vectors are concatenated and passed through a shared convolution to produce attention weights. This grouping strategy reduces the parameter count dramatically: GGCA uses only 0.03 thousand parameters compared to 0.40 thousand for standard coordinate attention, a reduction of more than 90 percent.
Cross-dimension interaction for richer context. Unlike approaches that treat each dimension independently, the GGCA module computes a joint representation that captures interactions between different spatial dimensions. When the model identifies a suspicious region along the depth axis, the cross-dimension component allows this signal to influence the height and width attention weights simultaneously, enabling more coherent 3D spatial reasoning.
Placement in the encoder. GGCA is applied at each resolution level of the V-Net encoder. This means it operates on features of decreasing spatial size but increasing semantic richness as the encoder deepens. By attending to cross-dimensional spatial context at every scale, the encoder learns progressively more abstract but spatially calibrated representations that distinguish nodule tissue from mimicking structures.
The decoder's unique challenge. During upsampling in the decoder, fine-grained boundary details must be reconstructed from coarser feature representations. Standard convolution treats all feature channels equally during this process. However, different feature channels encode different aspects of nodule morphology, from interior texture to edge sharpness, and selectively weighting these channels during reconstruction can improve boundary precision.
How split attention works. The GSA module divides input feature channels into multiple branches and applies separate convolutions with different receptive fields to each branch. The outputs are then fused using attention-weighted summation, where weights are determined by a compact attention mechanism that compares responses across branches. This multi-branch design allows the model to capture features at different scales simultaneously and adaptively select the most informative scale per channel.
Parameter efficiency over selective kernel networks. The standard selective kernel (SK) attention mechanism, which inspired GSA, uses 40.51 thousand parameters in this architecture. The GSA module achieves comparable multi-scale sensitivity with only 8.51 thousand parameters, a reduction of approximately 79 percent. This efficiency is achieved by grouping channels before the split operation and sharing parameters across groups.
Impact on boundary reconstruction. During ablation studies, inserting only the GSA module into the V-Net decoder improved the Dice Similarity Coefficient from 0.7564 to 0.7747 compared to the baseline. This improvement is particularly visible at nodule boundaries where smooth, accurate delineation requires integrating multi-scale decoder features rather than relying on a single receptive field.
The skip connection problem. Skip connections in V-Net directly concatenate encoder features with decoder features at corresponding resolution levels. This design assumes that all encoder features are equally useful for reconstruction, but in practice encoder features and decoder features carry different types of information. Encoder features have higher spatial detail but lower semantic specificity, while decoder features have higher semantic context but lower spatial resolution. Naively concatenating them without selection introduces noise that degrades segmentation precision.
Dual-input design for guided fusion. The DGA module takes both the encoder feature map and the decoder feature map as inputs simultaneously. It computes a channel-wise attention weight using a compact parallel pathway that processes both inputs together. The resulting attention map reweights the encoder features before they are fused with the decoder features, selectively suppressing encoder channels that are inconsistent with the semantic context provided by the decoder at that resolution level.
Why this helps with adherent nodules. For nodules attached to the pleural wall, encoder features often contain strong signals from both the nodule and the adjacent wall tissue, since the two appear similar in local texture. The DGA module uses the decoder's semantic context, which encodes the expected shape and location of the nodule, to suppress wall-related encoder channels and emphasize nodule-specific ones. This selective fusion is particularly effective at resolving ambiguous boundaries.
Ablation confirms DGA's strongest contribution. When only the DGA module was added to the baseline V-Net, it improved DSC from 0.7564 to 0.7842, the largest single-module improvement in the ablation study. This result demonstrates that refining skip connections is more impactful than improving the encoder or decoder alone, because skip connections directly determine how much boundary detail from the encoder is preserved in the final output.
Dataset and evaluation protocol. All experiments used the LUNA16 benchmark dataset, which contains 888 CT scans with 1,186 annotated pulmonary nodules spanning a diameter range of 3 to 30 mm. The dataset was split into 90 percent training and 10 percent testing. Performance was measured using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Precision, and Recall.
Improvement over baseline V-Net. The DGCA V-Net improved DSC from 0.7564 to 0.7921, a gain of 4.72 percent. IoU improved from 0.6231 to 0.6662, a gain of 6.92 percent. Precision increased from 0.7874 to 0.8102, and Recall improved from 0.7627 to 0.7993. These consistent gains across all four metrics confirm that the attention modules reduce both false positives and false negatives simultaneously.
State-of-the-art comparison. Among all models tested, DGCA V-Net achieved the highest DSC at 0.7921 and the highest Precision at 0.8102. TransUNet, a transformer-based baseline, scored 0.7875. nnU-Net, a widely used self-configuring segmentation framework, achieved 0.7911. WingsNet, another attention-augmented network, reached 0.7919. DGCA V-Net surpassed all of these without the computational overhead of full transformer architectures.
Computational efficiency maintained. Total training time for DGCA V-Net was 11,892 seconds, compared to 11,862 seconds for the baseline V-Net. This difference of 30 seconds across an entire training run demonstrates that the three attention modules add negligible computational cost, making the improvements purely additive in terms of accuracy without meaningful penalty in training time or inference speed.
Systematic ablation confirms each module's value. Four configurations were evaluated: baseline V-Net alone (DSC 0.7564), V-Net with GGCA only (DSC 0.7725), V-Net with GSA only (DSC 0.7747), V-Net with DGA only (DSC 0.7842), and the full DGCA V-Net with all three modules (DSC 0.7921). Each module independently improved performance, and the combination of all three produced further gains beyond any single module.
Complementary roles confirmed. The ablation results show that the three modules are not redundant. GGCA improves encoder feature discrimination (encoder-side gain), GSA improves decoder boundary reconstruction (decoder-side gain), and DGA improves how encoder and decoder information is fused at skip connections (fusion-side gain). Because each module targets a distinct architectural component, their contributions are additive rather than overlapping.
Qualitative improvements on hard cases. Visual comparisons in the paper show that the DGCA V-Net more accurately handles the two most difficult nodule types. For adherent nodules that are attached to vessel walls, the model produces sharper, more accurate boundaries compared to baseline V-Net which tends to merge the nodule with adjacent structures. For calcified nodules with high-density CT values, the baseline tends to under-segment due to the atypical intensity, while DGCA V-Net better delineates the full nodule extent.
Limitations acknowledged. The study was conducted on a single dataset, LUNA16, which contains nodules within a specific size range. Performance on very small nodules (under 3 mm diameter) or on nodules from different scanner protocols is not evaluated. Future work is needed to validate generalizability across different CT acquisition parameters and to extend the approach to detection tasks, not just segmentation of already-located nodules.