TopoGate: Quality-Aware Topology-Stabilized Gated Fusion for Longitudinal Low-Dose CT New-Lesion Prediction

TL;DR

TopoGate, a lightweight model that combines the follow-up appearance view with the subtraction view and controls their influence through a learned, quality-aware gate, is presented, which is simple, interpretable, and practical for reliable longitudinal LDCT triage.

Abstract

Longitudinal low-dose CT follow-ups vary in noise, reconstruction kernels, and registration quality. These differences destabilize subtraction images and can trigger false new lesion alarms. We present TopoGate, a lightweight model that combines the follow-up appearance view with the subtraction view and controls their influence through a learned, quality-aware gate. The gate is driven by three case-specific signals: CT appearance quality, registration consistency, and stability of anatomical topology measured with topological metrics. On the NLST--New-Lesion--LongCT cohort comprising 152 pairs from 122 patients, TopoGate improves discrimination and calibration over single-view baselines, achieving an area under the ROC curve of 0.65 with a standard deviation of 0.05 and a Brier score of 0.14. Removing corrupted or low-quality pairs, identified by the quality scores, further increases the area under the ROC curve from 0.62 to 0.68 and reduces the Brier score from 0.14 to 0.12. The gate responds predictably to degradation, placing more weight on appearance when noise grows, which mirrors radiologist practice. The approach is simple, interpretable, and practical for reliable longitudinal LDCT triage.

Figures (5)

• Figure 1: Longitudinal LDCT pair. (Left) Follow-up (FU); (Right) Registered baseline (BL_reg). Differences in reconstruction and residual misalignment can destabilize temporal subtraction $\Delta$, motivating quality-aware fusion.
• Figure 2: TopoGate framework. We deformably register baseline CT to the follow-up to obtain $\mathrm{BL_{reg}}$, then crop aligned 3D ROIs around each lesion point $\mathbf{p}^{(i)}$ and compute the temporal difference $\Delta=\mathrm{FU}-\mathrm{BL_{reg}}$. Two 3D encoders extract appearance and difference embeddings $(f_{\mathrm{app}}, f_{\Delta})$. A quality vector $\mathbf{q}=[q_{\text{ct}},q_{\text{reg}},q_{\text{topo}}]\in[0,1]^3$ controls a monotonic gate $\alpha=\sigma(w_1 q_{\text{ct}}+w_2 q_{\text{topo}}-w_3 q_{\text{reg}}+b)$, which adaptively fuses the two view-specific predictions: $s=\alpha\,g_{\mathrm{app}}(f_{\mathrm{app}})+(1-\alpha)\,g_{\Delta}(f_{\Delta})$, producing the calibrated output $\hat{y}=\sigma(s)$.
• Figure 3: Gate response vs. quality. Larger $\alpha$ (color) is associated with higher CT quality and topology stability, increasing trust in the appearance branch.
• Figure 4: Effect of quality filtering. Removing low-quality pairs increases AUROC and reduces Brier, improving reliability.
• Figure 5: Robustness. As simulated noise increases, the mean gate weight $\alpha$ increases monotonically, shifting weight from $\Delta$ to appearance.