Prototype code for chest X-ray image classification experiments using deep learning backbones, thoracic preprocessing and evaluation utilities.
The repository is kept as a technical portfolio example of medical imaging AI work: preprocessing imaging data, comparing model configurations, evaluating outputs and documenting the difference between experimental performance and clinical validation.
This project explores questions that commonly appear in medical AI development:
- How should imaging inputs be prepared before model training?
- Does masking the thoracic region change model behavior?
- How should class imbalance and validation splits be handled?
- What is needed before model results can become clinically meaningful evidence?
The code is not presented as a clinical product. It is a research and experimentation repository.
The project references chest X-ray data labelled for image classification experiments, including work connected with ACIL/BWH public research context.
Raw images are not included in this repository. Some scripts refer to local paths or previously generated datasets and model artifacts.
- Classification scripts.
- Preprocessing and image utility functions.
- Optional thoracic masking workflow.
- Evaluation and prediction utilities.
- Explainability-related scripts.
- Hyperparameter and training support code.
.
├── classification_inference.py
├── train_models.py
├── generate_explainability_maps.py
├── explainability/
├── evaluation/
├── image_processing/
└── training_utils/
The project includes:
- Grayscale conversion, resizing and normalization.
- Optional thoracic mask application.
- Transfer learning using deep learning backbones such as Xception, Inception-ResNet and EfficientNet-style models.
- Hyperparameter exploration.
- Evaluation through Keras metrics and custom prediction metrics.
- Grad-CAM style explainability utilities.
This repository is a research prototype. It has not been clinically validated and is not intended for diagnosis, triage, treatment decisions or patient care.
Known limitations:
- Raw data, trained weights and some generated artifacts are not included.
- Some scripts still assume local file structures from the original research environment.
- Results should be considered exploratory.
- Clinical use would require formal dataset governance, external validation, subgroup and bias analysis, calibration, workflow integration assessment and prospective evaluation.
This project demonstrates applied medical imaging AI work and validation-aware thinking. It is relevant to digital health, clinical validation and medical AI consulting because it shows how clinical imaging questions become preprocessing decisions, model choices and evaluation workflows.
MIT License.