ADNI-based Alzheimer’s Disease Decision Support Backend

This repository contains the backend for our senior design project: a Decision Support System for Alzheimer’s disease based on structural MRI. The service exposes a FastAPI endpoint that accepts 3D NIfTI brain scans, performs preprocessing (orientation and registration to MNI space), extracts a representative slice, and uses a deep learning classifier (InceptionV3 in PyTorch) to predict one of three labels:

• AD – Alzheimer’s Disease
• MCI – Mild Cognitive Impairment
• CN – Cognitively Normal

Important: This project is for research and educational purposes only and not approved for clinical use.

---

Project Structure

• code/api-script.py – FastAPI application defining the MRI classification API (/predict/).
• code/inference_only/ – Standalone inference scripts for different architectures (InceptionV3, ViT, ResNet18, EfficientNet-B3/B7, AlexNet, VGG16).
• code/training_without_balancing/, code/training_with_undersampling/, code/training_with_augmentation_updated/ – Training scripts and notebooks exploring various CNN/ViT models and data augmentation strategies.
• code/preprocessing_nifti/ – Scripts for NIfTI preprocessing (FSL orientation, FLIRT registration, etc.).
• code/skull-stripping/ – Skull-stripping scripts (e.g. SynthStrip, BET).
• code/augmentation_scripts/ – Data augmentation utilities for MRI slices.
• code/nii_to_png.py, code/view_nii.py, code/select-image.py, code/check_middle_slice.py – Utilities for converting and visualizing NIfTI/MRI data.
• code/model_outputs.md – Summary of training and inference performance across different architectures and augmentation settings.
• requirements.txt – Python backend dependencies.

The production-facing backend is centered on code/api-script.py; the rest of the code documents the model development and experimentation process.

---

Backend Architecture

• Framework: FastAPI
• Server: Uvicorn
• Deep Learning Library: PyTorch + Torchvision
• Primary Deployed Model: InceptionV3
  • Pretrained on ImageNet, fine-tuned for 3-way AD/MCI/CN classification
  • Model weights expected at ../InceptionV3_model.pth (relative to code/api-script.py)
• Input format: 3D NIfTI MRI volume (.nii or .nii.gz)
• Preprocessing pipeline (in api-script.py):
  1. Save uploaded NIfTI to ./uploaded_nifti/.
  2. Reorient to standard using FSL fslreorient2std → ./reoriented/.
  3. Linear registration to MNI152 template using FSL flirt → ./registered/.
  4. Load registered NIfTI with NiBabel and extract the middle axial slice.
  5. Intensity clipping (1st–99th percentile) and rescaling to ([0, 255]) with skimage.exposure.
  6. Convert to a grayscale PIL.Image, then to a 3‑channel tensor.
  7. Resize to (299 \times 299), normalize with ImageNet mean/std.
  8. Run a forward pass through InceptionV3 and apply softmax.

The /predict/ endpoint returns the predicted class and associated probability.

---

Requirements

Python & Libraries

• Python: 3.8+ (tested with modern 3.x)
• Python dependencies (from requirements.txt):
  • fastapi
  • uvicorn[standard]
  • torch
  • torchvision
  • Pillow
  • python-multipart

Install them with:

python -m venv .venv
.venv\Scripts\activate  # Windows PowerShell
pip install -r requirements.txt

System Dependencies (FSL)

The backend relies on FSL command-line tools for orientation and registration:

• fslreorient2std
• flirt
• MNI template: MNI152_T1_1mm_brain

In code/api-script.py the template is referenced as:

MNI_TEMPLATE = "/usr/share/fsl/data/standard/MNI152_T1_1mm_brain"

On your system, you may need to:

• Install FSL and ensure its binaries are in your PATH.
• Update the MNI_TEMPLATE path to match your local installation.

Model Weights

Place the trained InceptionV3 weights file at:

• InceptionV3_model.pth in the project root, so that from code/api-script.py it is reachable as:

model.load_state_dict(
    torch.load("../InceptionV3_model.pth", map_location=device, weights_only=False)
)

---

Running the API Server

From the project root:

.venv\Scripts\activate  # if not already active
python -m uvicorn code.api_script:app --reload --host 0.0.0.0 --port 8000

---

API Usage

Endpoint

• Method: POST
• URL: /predict/
• Content type: multipart/form-data
• Field: file – uploaded NIfTI file (.nii or .nii.gz)

The backend checks the content_type of the uploaded file and expects one of:

• application/nii
• application/nii.gz
• application/octet-stream

Example Request (Python)

import requests

url = "http://localhost:8000/predict/"
files = {"file": open("example_subject.nii.gz", "rb")}

response = requests.post(url, files=files)
print(response.json())

Example Response

{
  "predicted_class": "AD",
  "probability": 0.82
}

Error Responses

• 400 Bad Request – unsupported file type or invalid NIfTI file.
• 500 Internal Server Error – failures in FSL preprocessing (reorient/FLIRT) or inference.

---

Model Development & Performance

We experimented with a range of image classification architectures on MRI slices derived from ADNI-like T1-weighted images, including:

• AlexNet
• VGG16
• ResNet18 / ResNet50
• InceptionV3
• EfficientNet-B3 / B7
• Vision Transformer (ViT)
• MobileNet_V2
• DenseNet-121

Key details and metrics are summarized in code/model_outputs.md, including:

• Accuracy for different architectures on held-out sets.
• 5‑fold cross‑validation results for various augmentation schemes (e.g., horizontal/vertical flips, different dataset versions).
• Attempts at mitigating overfitting (early stopping, learning rate schedulers, modified fully-connected heads).
• Impact of preprocessing (skull stripping with SynthStrip/BET, orientation and registration, intensity standardization).

Overall, InceptionV3 with appropriate augmentation and hyperparameters emerged as a strong candidate for deployment in the backend.

---

Clinical Disclaimer & Limitations

• This system is not a medical device and has not been validated for clinical decision-making.
• Performance metrics in code/model_outputs.md are based on specific experimental splits and may not generalize to other cohorts or scanners.
• The model predicts AD/MCI/CN status from a single MRI acquisition and does not incorporate longitudinal, cognitive, genetic, or demographic information.
• Predictions should be interpreted as supporting information only, alongside expert clinical judgment and additional diagnostic tests.

---

Future Work

Potential extensions for this backend include:

• Exposing additional endpoints (e.g., returning Grad-CAM heatmaps to aid interpretability).
• Supporting subject-wise cross-validation and data leakage safeguards more explicitly in the pipeline.
• Integrating with a frontend dashboard for clinicians to upload scans, visualize slices, and review predictions over time.
• Logging predictions and metadata to a database for retrospective analysis.

---

Acknowledgments

This project was developed as part of a senior design project focused on supporting early detection and staging of Alzheimer’s disease using MRI data and deep learning. It builds upon:

• The ADNI initiative and similar neuroimaging datasets.
• Prior research on MRI-based classification of AD, MCI, and CN.
• The open-source ecosystems of PyTorch, FastAPI, FSL, and related tooling.