--- license: mit library_name: pytorch tags: - image-classification - explainable-ai - medical-imaging - grad-cam --- # T1 MLOps Stack — ChestXplain Checkpoint This repository contains the approved DenseNet121 checkpoint used by the ChestXplain research/demo application. The application performs multi-label classification of 14 thoracic conditions and provides Grad-CAM visual explanations. ## Training provenance - **Architecture:** DenseNet121 with a multi-label classification head - **Dataset:** NIH ChestX-ray14-derived training subset - **Training subset:** 20,000 images - **Checkpoint epoch:** 6 - **Validation AUC:** 0.813 (recorded in the checkpoint metadata) - **Checkpoint SHA-256:** `5bf5e15396805ac85a5c1f1839813bc4ea3a66ff694f5ed6f12c1fd2eaee0cbf` - **Artifact:** `densenet121_chestxray.pth` The checkpoint is provided as a model artifact for the associated portfolio MLOps demonstration. NIH images and source data are not included in this repository. ## Intended use Research, engineering evaluation, and educational/demo use only. This artifact is not validated for clinical deployment and must not be used to diagnose, treat, or make decisions about patients. ## Limitations - The model was trained on a 20,000-image subset rather than the complete NIH ChestX-ray14 dataset. - Dataset labels are noisy and may not represent definitive clinical diagnoses. - Performance may vary across institutions, scanners, patient populations, acquisition protocols, and disease prevalence. - Reported metrics are not evidence of clinical efficacy, safety, fairness, or regulatory approval. - Grad-CAM highlights model attribution regions; it is not a clinical explanation or proof of pathology. - The checkpoint has not undergone external validation, prospective evaluation, calibration analysis, or regulatory review. ## License and provenance The model artifact is released under the repository's MIT license for research/demo purposes, subject to the terms and attribution requirements of the underlying NIH ChestX-ray14 dataset and its original publication. See the associated source project for inference code, evaluation context, and full citations: