Temporal Prior-Guided Segmentation of Pulmonary Lesions in Longitudinal CT: Input-Level Versus Prompt-Level Prior Injection

Background/Objectives: In longitudinal computed tomography follow-up, lesion boundaries from the previous timepoint are routinely available but usually encoded as binary masks for automated segmentation. We convert the expert-drawn t − 1 reference mask into a continuous spatial prior through exponential distance encoding, without providing the previous image to the network, and examine whether the site at which this prior is injected matters more than its resolution or encoding. Methods: A total of 244 sequential t − 1 → t pairs from 136 patients at a single centre were analysed with patient-level five-fold cross-validation. The prior was injected into a network built on Segment Anything Model 2 either as a full-resolution fourth input channel or on the stride-16 grid via the pretrained mask-prompt encoder; a control arm reduced the channel prior to the same 32 × 32 grid, and a further arm was trained with the prior binarised. The four-channel and prior-free arms were replicated with EVA-02. Comparators were a copy-forward baseline and a two-channel nnU-Net. Results: The four-channel model reached a mean Dice of 0.739, significantly exceeding the prior-free model at 0.672, while copy-forward reached 0.410. In 27 pairs with no mask–target overlap, it still achieved 0.575, compared with 0.482 without the prior (p = 0.08). The prompt pathway reached 0.715, whereas the 32 × 32 control retained the full gain (0.743 vs. 0.739; 95% confidence interval [CI] of the difference −0.013 to +0.021) and exceeded the prompt pathway even in lesions spanning 1–2 stride-16 cells. An arm trained with a binary prior matched the soft prior, indicating that the prior acts as a coarse regional cue rather than a shape template. With last-epoch weights, as used for nnU-Net, the four-channel model matched nnU-Net in mean Dice at 0.704 but had fewer complete failures and no empty predictions, compared with 14. Weighted κ for categorical change classification did not improve significantly, at 0.555 versus 0.505. Conclusions: As a soft spatial prior, the routinely available previous mask improves segmentation accuracy and markedly reduces complete misses. Where the prior enters the network matters more than its spatial resolution or its soft encoding. Without gains in categorical change detection, the method currently suits automated pre-delineation and cohort-level analysis.