A lightweight 3D lung-nodule network stops collapsing to all-background: bidirectional feature interaction lifts Dice from 0.2302 to 0.5233
Synopsis
Light3D-NoduleNet is a lightweight pure 3D network that closes a loop between nodule detection and segmentation through bidirectional feature interaction, reaching a Dice of 0.5233 and a median HD95 of 6.00 mm on the LIDC-IDRI validation set with 5.24 M parameters and 178.3 G FLOPs. In ablation, three variants lacking that interaction all stalled at Dice 0.2302 after 100 epochs, while the full model reached 0.4240 in main training and 0.5233 after three-stage progressive fine-tuning. On a 15 W Jetson Orin NX it measured 3.0 FPS with 435 MB of VRAM.
Interpretation
Nodule detection and segmentation now share one pure 3D lightweight network, where the detection branch's center heatmap constrains where the segmentation branch looks instead of the two tasks running separate inference pipelines. Earlier joint frameworks mostly shared only backbone features, with task heads exchanging no explicit information, so detection and segmentation needed separate models and inference pipelines. On the LIDC-IDRI validation set the network reaches Dice 0.5233 and a median HD95 of 6.00 mm with 5.24 M parameters and 178.3 G FLOPs, against 0.3688 for a reproduced 3D U-Net and 0.4532 for V-Net under the same training and preprocessing.
The network factorizes the standard 3x3x3 kernel into an in-plane 1x3x3 branch and an inter-slice 3x1x1 branch, and adapts the channel-attention kernel size to the channel count. CT in-plane pixel spacing is about 0.5-0.8 mm while inter-slice spacing is about 1.0-2.5 mm, so an isotropic cubic kernel spends redundant computation along the slice direction. The asymmetric depthwise separable convolution cuts kernel parameters by about 56%, and the attention kernel size rises from 3 to 5 with channel count, switching at the third level with 128 channels.
Ablation shows bidirectional feature interaction is the condition that lets a lightweight 3D network escape the all-background trivial solution, not a small accuracy gain. The baseline, the variant adding asymmetric convolution, and the variant further adding channel attention each trained for 100 epochs under identical configurations, and Dice never exceeded 0.2302. The full model under the same main-training configuration reached 0.4240 at epoch 70 and 0.5233 after three-stage progressive fine-tuning, while the parameter-matched Baseline-Wide at 5.123 M without bidirectional interaction stalled at 0.2749 after 100 epochs.
Perspective
Results come from segmentation evaluation on the LIDC-IDRI validation set and end-to-end detection on 117 CT scans of the LUNA16 validation set, with 64x64x64 voxel inputs; edge measurements were taken on a 15 W Jetson Orin NX 8 GB.
The authors state that the convergence explanation for bidirectional interaction remains a retrospective reading of the ablation behavior and design intent, without an information-theoretic quantitative analysis; gradient flow and early per-task loss evolution remain to be verified. The detection branch is positioned as an auxiliary task, with LUNA16 end-to-end CPM at 0.3524, rising to 0.3774 under weighted fusion with a compact false-positive reduction module; only three nodules below 6 mm were present in the validation set, limiting that group's representativeness.