Project case study · 2026

Pneumonia X-ray Classifier

A reproducible computer-vision study comparing a custom CNN with established architectures across internal and external chest X-ray datasets.

Role
Machine Learning Engineer
Year
2026
Technologies
PythonPyTorchComputer VisionFastAPI
0.941ROC-AUC
94.62%Recall

Résumé-backed project-level results; architecture and dataset split not specified.

View project on GitHub

Problem

Pneumonia screening requires evaluation that accounts for class balance, external data behavior, model capacity, and operational recall.

Approach

The study trains a custom CNN, compares established architectures, selects thresholds on validation predictions, and tests internal and external datasets.

Architecture

dataset manifests → preprocessing → training and checkpoints → threshold selection → internal/external evaluation → complexity and result reports

Evaluation

Within this résumé-backed academic study—comparing a custom CNN with established architectures, selecting thresholds on validation predictions, and evaluating internal and external chest X-ray datasets—the reported project-level results are 0.941 ROC-AUC and 94.62% recall. The accepted evidence does not assign both headline metrics to a particular architecture or dataset split, so they should not be read as scores achieved across every internal and external evaluation.

Reproducibility

Reusable scripts cover training, evaluation, fine-tuning, dataset manifests, and model-complexity reporting so experiments can be repeated and compared through the same workflow.

Limitations

This is an academic prototype, not a medical device. Outcomes depend on dataset composition and external validation, and the results should not be interpreted as clinical performance.