COGENT explains lung cancer AI with 3D Gaussian counterfactuals

Explainability remains a critical bottleneck for deploying deep learning in medical settings—especially for 3D imaging like CT scans. Existing methods work at the voxel level, producing dense, hard-to-interpret attribution maps that ignore the structured representations now common in 3D scene modeling. COGENT (Counterfactual Gaussian Explanations) flips this by operating directly in the parameter space of Gaussian-based volumetric representations, a technique popularized in recent 3D rendering work.

Developed by a team from Polish research institutions, COGENT builds on MedGS (a Gaussian-based medical scene representation) and the Sybil lung cancer risk prediction model. Through a differentiable rendering pipeline, it optimizes selected Gaussian primitives based on gradients from the downstream predictor, identifying which components of the 3D scene most influence the model's decision. The result is a counterfactual—what would need to change to alter the diagnosis—that is sparse, spatially localized, and anatomically consistent. In tests on lung CT scans, COGENT outperformed conventional explainability methods in quantitative comparisons and was favorably reviewed by medical experts for clinical meaningfulness. This representation-space approach offers a new lens for interpreting volumetric deep learning models, moving beyond pixel-level heatmaps to structured, actionable explanations.