kanaadlimaye/Alzheimers-Disease-Classifier

Alzheimer Classification using ResNet-152

This model performs multi-class classification of Alzheimer's disease stages from MRI brain scans.
It uses ResNet152 pretrained on ImageNet as a feature extractor with a custom fully-connected classification head.

The following repo contains the TensorFlow-Keras model (.h5) version. The PyTorch version (.pt) is yet to be released.

The eval metrics on both the models is the same.

---

Source Code (GitHub Repository)

https://github.com/kanaad-lims/Alzheimer-Classification-using-PyTorch

---

Model Description

This model classifies MRI scans into four stages of Alzheimer's disease:

Class	Description
MD	Mild Demented
MoD	Moderate Demented
ND	Non Demented
VMD	Very Mild Demented

The architecture uses transfer learning with a pretrained ResNet152 backbone followed by custom dense layers for classification.

---

Model Architecture

Input Image (3 × H × W tensor)
        ↓
ResNet152 Backbone (ImageNet pretrained)
        ↓
Global Average Pooling
        ↓
  2048 feature vector
        ↓
  Batch Normalization
        ↓
      Dense(512)
        ↓
      Dense(256)
        ↓
     Dropout(0.5)
        ↓
      Dense(128)
        ↓
     Dropout(0.3)
        ↓
       Dense(64)
        ↓
    Dense(4) (class logits)

Softmax is implicitly applied during training through CrossEntropyLoss.

---

Training Details

Training Hardware

• Kaggle Notebooks
• NVIDIA Tesla T4 GPUs

Hyperparameters

Parameter	Value
Batch Size	32
Optimizer	Adam
Loss Function	CrossEntropyLoss
Backbone	ResNet152 (ImageNet pretrained)

The ResNet backbone weights were frozen, and only the custom classifier head was trained.

---

Dataset

Alzheimer's Classification Dataset
https://www.kaggle.com/datasets/kanaadlimaye/alzheimers-classification-dataset

The dataset consists of MRI brain images categorized into four stages of Alzheimer's disease.

Images were preprocessed using:

• Resizing to 224 × 224
• Horizontal and vertical flips
• Rotation augmentation (±15°)
• Normalization

---

Evaluation Results

Metric	Value
Training Accuracy	~97%
Validation Accuracy	~94%
Test Accuracy	~93%
Test Loss	~0.22

Classification Report

              precision    recall    f1-score    support

          MD     0.92       0.94       0.93        86
         MoD     1.00       0.80       0.89         5
          ND     0.94       0.96       0.95       319
         VMD     0.94       0.91       0.92       230

    accuracy                           0.94
   macro avg     0.95       0.90       0.92
weighted avg     0.94       0.94       0.94

---

Intended Use

The model is intended for:

• Academic research
• Machine learning experimentation
• Medical imaging classification studies

It is not intended for clinical diagnosis.

---

Limitations

• The dataset contains class imbalance, particularly for the Moderate Dementia class.
• The model was trained on a relatively small dataset.
• MRI scans from different scanners or hospitals may reduce performance.

---

Ethical Considerations

This model is intended strictly for research and educational purposes.
Medical decisions should never rely solely on automated predictions.

---

Citation

If you use this work in research, please cite the associated dataset and project.

@misc{alzheimers_resnet152_classifier,
  title  = {Alzheimer Stage Classification using ResNet152},
  author = {Kanaad Limaye},
  year   = {2026}
}