SY-001Objects outside the patient, gown snaps and ECG electrodes, drive Sybil's risk predictions

Bartlomiej Sobieski, Jakub Grzywaczewski, Karol Dobiczek, Mateusz Wójcik, Tomasz Bartczak, Patryk Szatkowski, Przemysław Bombiński, Matthew Tivnan, Przemyslaw Biecek

SourceAuditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions

Instead of only asking which pixels Sybil looks at, the authors edited the CT scans themselves, using a generative model to swap regions out for plausible healthy tissue, and watched what happened to the predicted cancer risk. Two of the regions driving predictions turned out to sit outside the patient altogether: metal snaps on a hospital gown pressed against the scanner table, and ECG electrodes stuck to the chest. In one benign case those electrodes accounted for half the predicted risk.

Evidence
interventional
Key metric
50% of predicted risk traced to ECG electrodes in one benign case
Model
Sybil
Concepts
Shortcut
Datasets
LUNA25 [eval], iLDCT [eval], NLST (National Lung Screening Trial) [train]
Methods
gSHNAP (generalized SHNAP) [primary], System-embedded diffusion bridges [primary], n-Shapley values [primary]
Related work
An overview of melanoma detection in dermoscopy images using image processing and machine learning [context]
Related findings
SY-003
Extraction
manual-extraction