Ashapu/Alzheimer_Classification_Model

🧠 Alzheimer's Disease MRI Multi-Stage Classifier (ResNet-18)

A high-sensitivity deep learning vision model for 4-stage Alzheimer's disease classification and 512-dimensional latent feature extraction from structural brain MRI scans.

⚠️ Clinical Research Disclaimer: This model is distributed strictly for scientific, academic research, and educational purposes. It is not an FDA/CE-cleared medical device and should never replace clinical evaluations, expert neurological consultations, or standard diagnostic workups (such as MMSE, MoCA, PET, or CSF biomarkers).

---

📌 Executive Summary

Early diagnosis of Alzheimer's disease (AD) is critical for timely therapeutic interventions, clinical trial enrollment, and personalized patient care. This model uses transfer learning via a customized ResNet-18 architecture fine-tuned on 86,437 brain MRI slices.

The model serves two primary modalities:

1. Multi-Stage Disease Classification: Categorizes structural MRI slices into 4 clinical progression tiers.
2. Latent Feature Extraction (Embedding Engine): Generates 512-dimensional semantic representations from the penultimate pooling layer for Content-Based Image Retrieval (CBIR) and visual case similarity matching.

---

🔬 Model Architecture & Design

Input: Brain MRI Scan (224 × 224 × 3 RGB)
   │
   ▼
┌────────────────────────────────────────────────────────┐
│               ResNet-18 Feature Backbone               │
│  (7x7 Conv, BatchNorm, MaxPool, Residual Blocks 1-4)   │
│                 11,176,512 Parameters                  │
└────────────────────────────────────────────────────────┘
   │
   ▼
┌────────────────────────────────────────────────────────┐
│             Adaptive Average Pooling Layer             │
│            Output: 512-dimensional vector              │
│       ★ Exportable for Case Similarity Search ★        │
└────────────────────────────────────────────────────────┘
   │
   ▼
┌────────────────────────────────────────────────────────┐
│             Fully Connected Classifier Head            │
│                 Linear(in: 512, out: 4)                │
└────────────────────────────────────────────────────────┘
   │
   ▼
Softmax Probabilities:
[Mild Dementia, Moderate Dementia, Non Demented, Very Mild Dementia]

Technical Specifications

• Base Architecture: torchvision.models.resnet18 (ImageNet-1K pre-trained backbone)
• Input Dimensions: 3 × 224 × 224 pixels (normalized to ImageNet mean/standard deviation)
• Parameters: 11,178,564 total parameters (~44.8 MB checkpoint size)
• Latent Vector Size: 512 dimensions (avgpool output)
• Optimization Strategy: Two-stage transfer learning with frozen residual stages and trainable classification head.

---

🏷️ Clinical Classification Categories

The model stratifies scans into four standard diagnostic classes:

Class ID	Diagnosis Label	Clinical Manifestation & Radiological Correlates
0	Mild Dementia	Noticeable cognitive impairment; moderate hippocampal and temporal lobe atrophy visible on structural T1-weighted images.
1	Moderate Dementia	Marked memory loss, impaired executive function; generalized cortical thinning, prominent sulcal widening, and ventricular enlargement.
2	Non Demented	Normal age-appropriate brain volume; preserved cortical thickness without localized atrophy.
3	Very Mild Dementia	Early Prodromal / Mild Cognitive Impairment (MCI); subtle medial temporal lobe volume changes often challenging for unassisted visual inspection.

---

📊 Comprehensive Performance & Benchmark Results

1. Training Progression

Trained across 15 epochs on NVIDIA Tesla T4 hardware using WeightedRandomSampler and CrossEntropyLoss(label_smoothing=0.1).

Epoch	Training Loss	Validation Accuracy	Status
1	0.8931	71.83%	Best checkpoint
3	0.7782	72.21%	Best checkpoint
5	0.7611	73.10%	Best checkpoint
6	0.7585	75.15%	Best checkpoint
9	0.7557	75.94%	Best checkpoint
11	0.7510	79.55%	🏆 Global Best Checkpoint
15	0.7355	76.68%	Converged

---

2. Validation Metrics (Held-out Set: 17,288 MRIs)

========================================================================
                   PRECISION    RECALL    F1-SCORE    SUPPORT (COUNT)
========================================================================
Mild Dementia         0.32       0.84       0.47           1,022
Moderate Dementia     0.27       0.97       0.42             121
Non Demented          0.94       0.81       0.87          13,390
Very Mild Dementia    0.55       0.53       0.54           2,755
------------------------------------------------------------------------
Accuracy                                    0.77          17,288
Macro Average         0.52       0.79       0.57          17,288
Weighted Average      0.83       0.77       0.79          17,288
========================================================================

3. Confusion Matrix Breakdown

                       PREDICTED
                Mild    Mod    Non    VeryMild   │ Total (Actual)
 ACTUAL       ┌───────┬──────┬───────┬──────────┐│
 Mild         │  855  │  27  │   85  │    55    ││  1,022
 Moderate     │    3  │ 117  │    0  │     1    ││    121
 Non Demented │ 1,220 │ 199  │10,818 │ 1,153    ││ 13,390
 Very Mild    │  562  │  92  │  635  │ 1,466    ││  2,755
              └───────┴──────┴───────┴──────────┘│

4. Key Strengths in Clinical Screening:

• Exceptionally High Sensitivity (Recall) for Severe Stages:
  • Moderate Dementia Recall: 96.69% (117 out of 121 cases correctly detected, only 4 missed).
  • Mild Dementia Recall: 83.66% (855 out of 1,022 detected).
• High Specificity for Healthy Scans:
  • Non Demented Precision: 93.76% (When the model identifies a scan as healthy, it has ~94% reliability).
• Weighted Balance:
  • High weighted average precision (83%) and F1-score (0.79) over the entire 17,288 evaluation corpus.

---

💻 How to Use

Installation

pip install torch torchvision pillow requests

1. Direct Inference Pipeline (Classification)

import torch
import torch.nn as nn
from torchvision import models, transforms
from PIL import Image

# Setup device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Initialize architecture matching the trained model
model = models.resnet18(weights=None)
model.fc = nn.Linear(model.fc.in_features, 4)

# Load checkpoint
model_path = "resnet18_alzheimer_best.pth"
model.load_state_dict(torch.load(model_path, map_location=device))
model.to(device)
model.eval()

# Preprocessing transforms (ImageNet standard)
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

classes = [
    "Mild Dementia",
    "Moderate Dementia",
    "Non Demented",
    "Very Mild Dementia"
]

# Run prediction
img = Image.open("sample_brain_mri.png").convert("RGB")
tensor = transform(img).unsqueeze(0).to(device)

with torch.no_grad():
    logits = model(tensor)
    probs = torch.softmax(logits, dim=1)[0].cpu().numpy()
    pred_idx = torch.argmax(logits, dim=1).item()

print(f"Top Prediction: {classes[pred_idx]} ({probs[pred_idx]*100:.2f}%)")
print("\nClass Probabilities:")
for label, p in zip(classes, probs):
    print(f"  • {label:20s}: {p*100:.2f}%")

---

2. Feature Extraction for Case Retrieval (Latent Embeddings)

# Strip the final classification layer to extract 512-d feature vectors
feature_extractor = nn.Sequential(*list(model.children())[:-1])
feature_extractor.eval()

with torch.no_grad():
    # Returns shape: (1, 512)
    embedding = feature_extractor(tensor).squeeze().cpu().numpy()

print(f"Extracted embedding dimension: {embedding.shape}")
# You can now compute cosine similarity against a database of MRI scans!

---

⚙️ Training Protocol & Reproducibility

Hyperparameter	Configuration
Dataset Size	86,437 total images
Train / Validation Split	80% (69,149 images) / 20% (17,288 images), seeded at 42
Batch Size	32 (optimized for GPU memory throughput)
Optimizer	AdamW(lr=1e-3, weight_decay=1e-4)
Learning Rate Schedule	ReduceLROnPlateau(mode='max', patience=2, factor=0.5)
Loss Function	CrossEntropyLoss(label_smoothing=0.1)
Sampling Technique	WeightedRandomSampler (inverse class frequency balancing)
Data Augmentation	RandomRotation(degrees=10), RandomHorizontalFlip(p=0.5)

---

🛡️ Ethical Considerations & Bias Statement

• Class Distribution: In typical clinical cohorts, the ratio of healthy elderly individuals exceeds diagnosed dementia patients. The training dataset reflects this real-world imbalance, which was mitigated via WeightedRandomSampler to boost minority-class detection sensitivity.
• Slice Bias: 2D slice analysis cannot fully capture volumetric brain changes (e.g., total hippocampal volume) as effectively as complete 3D volumetric MRI (e.g., NIfTI/DICOM T1 MPRAGE sequences).
• Intended Use: Screening assistance, radiologist second-reading triage, and educational visualization.

---

📜 Citation & Credits

If you use this model or its weights in your research or application, please cite:

@misc{alzheimers_mri_resnet18,
  title={Alzheimer's Disease MRI Multi-Stage Classifier and Feature Extractor using ResNet-18},
  author={Alzheimer's Early Detection AI Team},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co}}
}

---

Made with ❤️ for neuroscience and early Alzheimer's awareness.