FIDELIS — Reliability Estimators

Model artifacts for FIDELIS, an open-source tool that evaluates the calibration and robustness of trained deep-learning models in cancer imaging. FIDELIS operates post-hoc on a model's outputs — no retraining, no access to the original training data — and returns a clinician-facing reliability report.

• Code: https://github.com/ZhixiangWang-CN/FIDELIS
• Program: NCI ITCR (RFA-CA-27-019, U01)

Components (to be released)

• FIDELIS-Calibrate — post-hoc calibration assessment (Anatomy-Aware ECE / ΔA-ECE) and recalibration maps.
• FIDELIS-Robust — stability under simulated scanner / dose / protocol variation.
• Clinical Decision Impact (CDI) — maps reliability failures to potential changes in a clinical decision.

Status

Repository reserved during early development; trained components and full model cards will be uploaded and version-tagged with each release. Not a diagnostic device; not for autonomous clinical decision-making.

License

Code components: Apache-2.0. Standalone model weights: CC-BY-4.0.