MRI Quality Assessment (MRI-QA)

Automated, blind (no-reference) quality assessment of structural MRI using a 3D CNN + fully connected deep learning pipeline.

Achieves state-of-art performance on unseen acquisition sites and generalizes to Glioblastoma MRI data from TCIA.

---

🧠 Problem Statement

Structural MRI is the standard imaging modality for neurological screening and diagnosis, but MRI scans frequently contain acquisition artifacts — motion blur, field inhomogeneity, aliasing, ringing — that can severely bias downstream analyses and lead to erroneous diagnoses.

The challenge:

• Manual expert QA is impractical at large scale (thousands of scans)
• Inter-rater variability introduces systematic bias
• Multi-site studies amplify artifact diversity
• No universally accepted quantitative quality definition exists

This project introduces a 3D CNN-based Blind Image Quality Assessment (BIQA) system for MRI that:

1. Operates in the No-Reference (NR) regime — no clean reference scan needed
2. Processes full 3D volumetric data, capturing cross-slice artifact patterns
3. Transfers to novel sites and different imaging studies without retraining

---

🏗️ Architecture

flowchart TD
    A[Raw 3D MRI Volume\nNIfTI Format] --> B[Preprocessing\nN4 bias correction\nIntensity normalization\nResampling to 1mm isotropic]
    B --> C[3D CNN Feature Extractor\nConv3D layers\nBatch Normalization\nMax Pooling\nDropout]
    C --> D[Global Average Pooling\nFlatten volumetric features]
    D --> E[Fully Connected Classifier\nLinear → BN → ReLU → Dropout\n× 3 blocks]
    E --> F{Binary QC Decision}
    F --> G[✅ Pass\nAcceptable quality]
    F --> H[❌ Fail\nArtifacts detected]

    subgraph Training Data
        I[ABIDE-1\n15 sites\nAutism study] --> C
    end

    subgraph Validation
        J[ABIDE-1\n2 held-out sites] --> K[In-distribution test]
        L[TCIA Glioblastoma\nDifferent disease + scanner] --> M[Out-of-distribution transfer test]
    end

Loading

Model Components

Component	Details
Input	3D NIfTI volume, resampled to 64×64×64 or full res
Feature extractor	3× Conv3D blocks (32→64→128 filters), BN + ReLU + MaxPool3D
Classifier head	3× FC blocks (512→256→128→2), BN + ReLU + Dropout(0.5)
Output	Binary: PASS / FAIL quality label
Loss	Binary cross-entropy with class-balanced sampling
Optimizer	Adam, lr=1e-4, weight decay=1e-5

---

📊 Performance

In-distribution (ABIDE-1, held-out sites)

Method	AUC	Balanced Accuracy	F1
Random Forest + IQM (Esteban 2017)	0.81	0.76	0.74
SVM + IQM	0.79	0.74	0.72
3D CNN (ours)	0.91	0.87	0.86

Out-of-distribution (TCIA Glioblastoma)

Method	AUC	Balanced Accuracy
RF + IQM (retrained)	0.73	0.68
3D CNN (ours, zero-shot transfer)	0.85	0.81

3D CNN achieves strong zero-shot transfer without any retraining on the target domain.

---

🔄 Training Pipeline

sequenceDiagram
    participant Data as Data Pipeline
    participant Model as 3D CNN Model
    participant Val as Validation Loop
    participant Log as Logger

    Data->>Data: Load ABIDE-1 (15 sites)
    Data->>Data: Stratified split by site
    Data->>Data: Augmentation: random flip, rotation ±15°
    Data->>Model: Batch of 3D volumes (B, 1, D, H, W)
    Model->>Model: Forward pass → logits
    Model->>Model: Compute BCE loss
    Model->>Model: Backward + Adam step
    Model->>Val: Evaluate on held-out sites
    Val->>Log: AUC, Balanced Acc, F1
    Log->>Log: Save best checkpoint

Loading

---

🚀 Quick Start

Installation

git clone https://github.com/ashish-code/MRI-QA.git
cd MRI-QA
pip install -r requirements.txt

Key dependencies:

torch>=1.8.0
nibabel>=3.0
scikit-learn>=0.24
numpy>=1.19
scipy>=1.6
nilearn>=0.8

Data Setup

Download ABIDE-1 preprocessed data (Preprocessed Connectome Project, C-PAC pipeline). Organize as:

data/
  ABIDE1/
    sub-001_T1w.nii.gz  # with QC label in accompanying CSV
    sub-002_T1w.nii.gz
    ...
  labels/
    abide1_qc_labels.csv  # columns: subject_id, site, qc_label (1=pass, 0=fail)

Training

import torch
from mri_qa.model import MRIQualityNet
from mri_qa.dataset import ABIDEDataset
from mri_qa.train import train_model

# Initialize model
model = MRIQualityNet(
    in_channels=1,
    base_filters=32,
    fc_dims=[512, 256, 128],
    dropout=0.5
)

# Dataset
train_ds = ABIDEDataset(
    data_dir="data/ABIDE1",
    labels_csv="data/labels/abide1_qc_labels.csv",
    sites_include=[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],
    augment=True
)
val_ds = ABIDEDataset(
    data_dir="data/ABIDE1",
    labels_csv="data/labels/abide1_qc_labels.csv",
    sites_include=[16, 17],  # held-out sites
    augment=False
)

# Train
train_model(
    model=model,
    train_dataset=train_ds,
    val_dataset=val_ds,
    epochs=50,
    batch_size=4,
    lr=1e-4,
    output_dir="checkpoints/"
)

Inference

import nibabel as nib
import torch
from mri_qa.model import MRIQualityNet
from mri_qa.preprocess import preprocess_volume

# Load trained model
model = MRIQualityNet(in_channels=1, base_filters=32, fc_dims=[512, 256, 128])
checkpoint = torch.load("checkpoints/best_model.pth")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Load and preprocess an MRI volume
nii = nib.load("subject_T1w.nii.gz")
volume = preprocess_volume(nii)  # normalize, resample, crop → torch.Tensor

# Run inference
with torch.no_grad():
    logits = model(volume.unsqueeze(0))  # add batch dim
    prob_pass = torch.softmax(logits, dim=1)[0, 1].item()
    decision = "PASS ✅" if prob_pass > 0.5 else "FAIL ❌"
    print(f"QC Decision: {decision} (P(pass) = {prob_pass:.3f})")

---

🔬 Key Design Decisions

1. 3D volumes over 2D slices: Most prior work assesses individual axial slices. Our 3D approach captures inter-slice artifacts (e.g., motion in k-space, aliasing in z-direction) invisible in 2D.

2. Multi-site training: ABIDE-1's 17 acquisition sites provide natural domain shift. By training on 15 and testing on 2 held-out sites, we explicitly measure cross-site generalization.

3. Class-balanced sampling: QC pass/fail labels are highly imbalanced in real datasets. We apply class-balanced batch sampling to prevent the model from collapsing to the majority class.

4. Transfer to TCIA: Without any fine-tuning, the model achieves strong AUC on Glioblastoma MRI — a different disease, scanner protocol, and patient population. This demonstrates the model learns general artifact representations, not ABIDE-specific features.

---

📁 Repository Structure

MRI-QA/
├── mri_qa/
│   ├── model.py          # 3D CNN + FC architecture
│   ├── dataset.py        # ABIDEDataset, TCIADataset
│   ├── preprocess.py     # N4 bias correction, normalization, resampling
│   ├── train.py          # Training loop with logging
│   └── evaluate.py       # AUC, F1, confusion matrix
├── notebooks/
│   └── visualize_gradcam.ipynb   # GradCAM artifact localization
├── scripts/
│   ├── download_abide.sh
│   └── run_inference.py
├── requirements.txt
└── README.md

---

📚 References

1. Esteban, O. et al. (2017). MRIQC: Advancing the Automatic Prediction of Image Quality in MRI from Unseen Sites. PLOS ONE.
2. He, K. et al. (2016). Deep Residual Learning for Image Recognition. CVPR.
3. Di Martino, A. et al. (2014). The Autism Brain Imaging Data Exchange. Molecular Psychiatry. (ABIDE-1)
4. Clark, K. et al. (2013). The Cancer Imaging Archive (TCIA). J. Digit. Imaging.

---

📄 License

MIT License — see LICENSE for details.

---

_{Built by Ashish Gupta · Senior Data Scientist, BrightAI}