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

Seungik Cho has introduced TopoGate, a novel AI architecture designed to solve a critical problem in lung cancer screening: false positive alarms in longitudinal low-dose CT (LDCT) follow-ups. These scans, used to monitor patients over time, often vary in quality due to differences in noise, reconstruction kernels, and registration, which destabilizes the 'subtraction images' (differences between scans) and triggers false alarms for new lesions.

TopoGate's innovation is a lightweight, interpretable model that fuses two data views: the raw follow-up CT appearance and the subtraction image. A learned 'gate' dynamically controls the influence of each view based on three quality signals: CT appearance quality, registration consistency, and the stability of anatomical topology measured via topological data analysis metrics. This allows the model to behave like a radiologist, placing more weight on the clearer appearance view when noise corrupts the subtraction view. Technically, on the NLST–New-Lesion–LongCT cohort of 152 scan pairs, TopoGate achieved an area under the ROC curve (AUC) of 0.65 (±0.05 SD) and a Brier score of 0.14, outperforming single-view baselines.

The practical impact is significant for clinical triage. When TopoGate's internal quality scores were used to filter out corrupted or low-quality scan pairs, performance improved markedly—AUC increased from 0.62 to 0.68 and the Brier score dropped from 0.14 to 0.12. This represents a potential 43% reduction in the relative increase of the AUC, directly translating to fewer unnecessary patient recalls and biopsies. The approach is described as simple and practical, offering a reliable, automated triage tool that can be integrated into existing screening workflows to make longitudinal LDCT analysis more robust and trustworthy.