Cethraian-Shift: Cross-Dataset Reliability, Calibration, and Explainability of Multi-Label Chest X-Ray Classification

# Cethraian-Shift Cethraian-Shift is a reproducible medical imaging research project investigating the reliability of multi-label chest X-ray classification across public datasets. The study evaluates how model performance, calibration, explainability, and prediction behavior transfer between NIH ChestX-ray14 and VinDr-CXR using a frozen, reproducible evaluation protocol. The project focuses on methodological transparency rather than proposing a new neural-network architecture. Analyses include cross-dataset discrimination, probability calibration, label-policy sensitivity, Grad-CAM++ localization, explanation-map stability, subgroup analyses, and deterministic failure auditing. This record contains the release manuscript, supplementary material, and the corresponding public repository snapshot used for the v1.0.0 release. ## Contents - Main manuscript- Supplementary material- Public source code- Reproducibility documentation- Evaluation results- Configuration files ## Study scope - DenseNet-121 evaluated under a frozen three-seed protocol- Cross-dataset evaluation on NIH ChestXray14 and VinDr-CXR- Calibration analysis- Label-policy sensitivity analysis- Grad-CAM++ explainability analysis- Cross-seed explainability stability- Deterministic and reproducible evaluation ## Limitations This work is a retrospective study using publicly available datasets. It does not introduce a novel neural-network architecture, does not claim clinical deployment readiness, and does not claim radiologist-level diagnostic performance.