DenseNet-201 Brain Stroke Classifier

I have a collection of brain scans—specifically MRI and CT images—saved in JPEG/PNG format. I need a deep-learning solution that can reliably detect the presence of stroke lesions using the DenseNet-201 architecture.

What I will hand over:
• MRI images (JPEG/PNG)
• CT images (JPEG/PNG)

What I need back:
• Clean, well-commented Python code (TensorFlow/Keras or PyTorch) that builds, trains, and validates a DenseNet-201 model on the supplied data
• All preprocessing and augmentation steps included in the pipeline so I can reproduce results
• A trained weight file plus an easy-to-run inference script or notebook that takes a new scan and outputs the stroke / no-stroke prediction probability
• Performance report covering accuracy, sensitivity, specificity, and ROC-AUC, along with confusion matrix visuals
• Optional but valuable: Grad-CAM or similar heat-map visualizations that highlight stroke-related regions for model interpretability

Please structure the project so I can drop the raw images into a folder, adjust a path in the config, and re-train or fine-tune if new data arrives.