SY-003Sybil processes pulmonary nodules almost additively, with limited pairwise interactions
Bartlomiej Sobieski, Jakub Grzywaczewski, Karol Dobiczek, Mateusz Wójcik, Tomasz Bartczak, Patryk Szatkowski, Przemysław Bombiński, Matthew Tivnan, Przemyslaw Biecek
Using SHNAP, the authors replaced each detected nodule with generated healthy tissue, constructed a nodule-free baseline, restored the original nodules in every possible coalition, and evaluated Sybil on each coalition. Across the full LUNA25 test split and iLDCT, main effects reproduced Sybil's outputs with local-fidelity R2 near 1 for the majority of cases, while adding pairwise interactions almost entirely eliminated the outlier tail.
Evidence
interventional
Key metric
main effects alone reproduce the model's outputs at local fidelity R2 near 1 for the majority of cases, and adding pairwise interactions removes almost all of the outlier tail
Caveat
The analysis relies on partially synthetic nodule removals, which the authors identify as the study's primary limitation despite a blinded radiologist study. The remaining low-fidelity cases typically involve rare, anomalously large nodules for which SDB struggles with reconstruction.