Revisiting Dual-Encoder Fusion and Optimal Transport Regularization in PET-CT Tumor Segmentation: A Matched-Control Study under Extended Training

Dual-encoder fusion and explicit cross-modal alignment are two common strategies for multimodal PET-CT segmentation, but their relative contributions remain unclear when training duration and model capacity are controlled. We therefore reassess dual-encoder fusion and Optimal Transport (OT) regularization in DpDNet using an extended 200-epoch protocol with convergence monitoring and five-fold cross-validation on 75 AutoPET-derived PET-CT cases. The dual encoder shows no significant Dice advantage over a parameter-matched, approximately compute-matched weight-shared two-stream encoder (\(\Delta\text{Dice} = 0.0076\), \(p = 0.61\)), but achieves higher Dice than a parameter-matched early-fusion encoder (\(\Delta\text{Dice} = 0.0465\), \(p = 0.016\)). Because the latter requires \(1.24\times\) the FLOPs, this difference cannot be attributed uniquely to fusion position. OT yields no significant late-training Dice improvement on either the dual encoder (\(\Delta\text{Dice} = 0.0091\), \(p = 0.58\)) or the weight-shared two-stream encoder (\(\Delta\text{Dice} = 0.0080\), \(p = 0.45\)), and a 20-epoch pilot advantage does not persist under the 200-epoch protocol. The previously observed pre-training–OT interaction is likewise not reproduced and remains exploratory. Voxel Dice does not distinguish the matched-capacity configurations, whereas lesion-level detection shows only nominal differences that do not survive multiple-comparison correction. These results indicate that conclusions about multimodal PET-CT segmentation depend strongly on control design, training duration, and evaluation metric.