Spaces:
Sleeping
π« Precision Diagnostics: AI-Powered Chest X-Ray Analysis & Explainability Engine
An end-to-end MLOps pipeline designed to analyze chest X-ray images for pediatric pneumonia detection with high clinical sensitivity. The system features a deep learning classification engine, an automated hyperparameter optimization suite, a FastAPI backend, and a React frontend dashboard. To assist clinical workflows, it integrates Grad-CAM (Gradient-weighted Class Activation Mapping) to visualize precisely where the model focuses its diagnostic attention.
ποΈ System Architecture & Model Workflow
The architecture is split into three decoupled components to maintain high performance, modularity, and easy software updates:
- The Core AI Brain: A fine-tuned
DenseNet121model trained with class-balanced cross-entropy loss and hyperparameter-tuned via Optuna. - The Inference Backend (FastAPI): Exposes a REST API (
/predict) that automatically downloads optimized model weights from cloud storage on startup, processes incoming images, and generates explainability layers. - The User Interface (React): A clinical dashboard for uploading X-rays, viewing diagnosis metrics, and rendering interactive Grad-CAM heatmaps.
π¦ Model Weights Storage (Large File Notice)
Deep Learning model binaries (.pth files) exceed standard Git tracking limits (>25MB). To follow MLOps best practices, model weights are hosted externally and are decoupled from the code repository.
The baseline model weights (v1.0.0) can be retrieved through two separate cloud channels:
| Platform | Repository / Resource URL | Primary Use Case |
|---|---|---|
| Hugging Face Hub | mobadara/pneumonia-densenet121 |
Production & Automated FastAPI Startup (Recommended) |
| GitHub Releases | Precision Diagnostics v1.0.0 Release | Manual Local Testing & Containment Backups |
Model Metrics Snapshot (v1.0.0)
- Architecture: DenseNet121 with custom 2-class fully connected classification head.
- Global Accuracy: 86.4% on a fully unseen test dataset.
- Pneumonia Recall (Sensitivity): 94.6% (Successfully flagged 369 out of 390 positive pathology samples). The training loss incorporates dynamic weights to minimize critical False Negatives.