Reproducibility package for “Predict–Compare–Correct: Closed-Loop Self-Correction for Future Lesion Change Prediction in Longitudinal Glioma Imaging”

This repository provides the code and reproducibility package for a closed-loop self-correcting framework for future lesion change prediction in longitudinal glioma imaging. The core method follows a predict-compare-correct pipeline: the model first predicts future lesion change, then compares this prediction against the corresponding future change target, constructs a correction signal, and feeds that signal back into a refinement path. The final mainline model included in this release is R_final, which achieved the best overall balance across clean performance, shrinkage-sensitive behavior, and stability.

The package includes release-ready training and evaluation scripts, configuration files, data preparation utilities, robustness evaluation scripts, documentation, and key result tables, figures, and reports. It also documents the main experiment lineage, including the reference model (A_before), transition version (E), mechanism probe (Rplus), and bridge probe (R-Bridge), so that readers can understand how the final model was derived and how different design choices affected behavior.

This upload is intended to support reviewer inspection and independent reproduction of the reported experiments. It does not contain the original medical imaging dataset itself. Users must prepare their own data paths and comply with the license and access conditions of the underlying source dataset. The repository documentation explains the expected data structure, manifest generation, training workflow, clean evaluation, and robustness evaluation under controlled perturbations.