ML Pipeline for Medical Imaging (VGG16 + MLOps)

Problem Statement

Chest cancer classification from medical images needs to be both accurate and reproducible — a model that can't be retrained and verified the same way twice isn't trustworthy in a healthcare context.

Solution

Fine-tuned VGG16 (pretrained on ImageNet) for chest cancer classification, achieving 75% accuracy, with the full training and validation workflow made reproducible and automated.

Architecture

VGG16 feature extractor fine-tuned for classification; DVC handles data versioning, MLflow tracks experiments, and the trained model is containerized with Docker and deployed to AWS (EC2, S3).

Results

• 75% classification accuracy using VGG16 transfer learning
• Fully reproducible, automated training and validation workflow via DVC + MLflow