README.md

metadata

title: 'CerebAI: AI-Powered Stroke Detection System'
emoji: 🧠
colorFrom: red
colorTo: indigo
sdk: streamlit
app_file: cerebAI.py
license: mit
sdk_version: 1.50.0

CerebAI: AI-Powered Stroke Detection System

Project Overview

CerebAI is a deep learning application designed to assist medical professionals by rapidly classifying CT scan images for the presence and type of stroke. Built on the advanced ConvNeXt architecture, the system provides a robust diagnosis coupled with a critical eXplainable AI (XAI) feature, ensuring predictions are transparent and medically intuitive. This project showcases high-performance multiclass classification and deployment readiness.

Key Technical Achievements

Metric	Score (Test Set)	Implication
Mean IoU (mIoU)	~0.9843	Top-tier performance for pixel-level prediction quality.
Test F1 Score (Weighted)	~0.9805	Excellent balance between Precision and Recall across all three classes.
Model Architecture	ConvNeXt Base	State-of-the-art model designed for robust feature extraction from medical images.

Interpretability (XAI Feature)

The system uses Integrated Gradients (IG) from the Captum library to generate a heatmap overlay.

• Function: IG highlights the specific pixels that most strongly influence the model's final diagnosis.
• Clinical Value: This visual evidence helps doctors verify the prediction by confirming the model is focusing on the actual pathology (the stroke region) and not on noise or artifacts.

Deployment and Setup

Local Run Instructions

1. Clone the Repository:

   git clone [https://github.com/ar-shenoy/cerebAI.git](https://github.com/ar-shenoy/cerebAI.git)
   cd cerebai_streamlit
   

2. Activate Environment: Ensure your virtual environment is active.

   .\venv\Scripts\activate
   

3. Install Dependencies:

   pip install -r requirements.txt
   

4. Place Model Weights: Ensure your trained model file (best_model.pth) is in the project root directory.
5. Launch App:

   streamlit run cerebAI.py
   

Streamlit Deployment (Cloud)

This repository is configured for one-click deployment on the Streamlit Community Cloud. The app is optimized to run on a shared CPU by limiting the Integrated Gradients calculation to 20 steps to ensure fast performance.