This dataset is an ML-ready derivative of the OASIS (Open Access Series of Imaging Studies) brain MRI database, curated and preprocessed for Alzheimer's disease stage classification. It contains 10,880 MRI brain scan images split across training, validation, and test sets, each labeled with one of four dementia severity stages.

The OASIS project is a collaborative initiative aimed at making brain MRI datasets freely available to the scientific community to accelerate discoveries in basic and clinical neuroscience. It is made available by:

Washington University Alzheimer's Disease Research Center
Dr. Randy Buckner, Howard Hughes Medical Institute (HHMI) at Harvard University
Neuroinformatics Research Group (NRG), Washington University School of Medicine
Biomedical Informatics Research Network (BIRN)

The images in this dataset span cross-sectional MRI scans from young, middle-aged, nondemented, and demented older adults, making it suitable for building and benchmarking automated Alzheimer's detection and staging models.

Supported Tasks

Task	Type
Alzheimer's Stage Classification	Multi-class Image Classification (4 classes)

Languages

Not applicable — this is an image dataset. All metadata and CSV files are in English.

Dataset Structure

Class Labels

The dataset uses four class labels representing stages of Alzheimer's disease progression:

Label	Abbreviation	Description
`NonDemented`	ND	No signs of cognitive impairment
`VeryMildDemented`	VMD	Very early-stage; subtle structural changes
`MildDemented`	MD	Mild cognitive decline; early neurodegeneration
`ModerateDemented`	MoD	Moderate cognitive decline; significant brain atrophy

Data Splits

Split	# Images	Files
Train	8,960	`Train/` + `CSV_datafiles/train.csv`
Validation	1,280	`Valid/` + `CSV_datafiles/valid.csv`
Test	640	`Test/` + `CSV_datafiles/test.csv`
Total	10,880

ℹ️ The training split has been augmented (see Preprocessing & Augmentation), yielding 2 augmented outputs per original training example.

Repository Structure

alzheimers-classification-dataset/
│
├── CSV_datafiles/
│   ├── train.csv        # Image paths + labels for training
│   ├── valid.csv        # Image paths + labels for validation
│   └── test.csv         # Image paths + labels for testing
│
├── Train/               # 8,960 augmented .png MRI images
├── Valid/               # 1,280 .png MRI images
└── Test/                # 640 .png MRI images

Data Fields

Field	Type	Description
`image`	`PIL.Image` / `.png` file	Preprocessed MRI brain scan (224×224 px)
`label`	`ClassLabel` (int)	Alzheimer's stage: 0=NonDemented, 1=VeryMildDemented, 2=MildDemented, 3=ModerateDemented
`filename`	`string`	Path to the image file (referenced in CSV files)

Data Instances

Example entry from a CSV file:

filename,label
Train/NonDemented/OAS1_0001_MR1_mpr_n4_anon_111_t88_gfc_tra_90.png,NonDemented
Train/VeryMildDemented/OAS1_0003_MR1_mpr_n3_anon_111_t88_gfc_tra_90.png,VeryMildDemented

Dataset Creation

Source Data

The raw MRI data originates from the OASIS Cross-Sectional MRI Database, which includes brain scans from:

Young and middle-aged adults (ages 18–59)
Nondemented older adults
Demented older adults (including individuals diagnosed with Alzheimer's disease)

Original OASIS scans were acquired using a 1.5T Vision scanner (Siemens) with a standardized MP-RAGE sequence. Each subject underwent multiple T1-weighted MRI acquisitions per session.

Curation & Processing by Dataset Author

The dataset author performed the following transformations on the raw OASIS data to make it ML-ready:

Format conversion: All MRI images converted to .png format.
Dataset restructuring: Images organized into labeled class folders.
CSV split files created: train.csv, valid.csv, and test.csv were generated to define reproducible dataset splits.
Preprocessing & augmentation applied (see below).

Annotations

Class labels (dementia stages) are derived from clinical assessments and CDR (Clinical Dementia Rating) scores provided in the original OASIS dataset:

CDR Score	OASIS Clinical Rating	Dataset Label
0	No dementia	`NonDemented`
0.5	Very mild dementia	`VeryMildDemented`
1.0	Mild dementia	`MildDemented`
2.0	Moderate dementia	`ModerateDemented`

Preprocessing & Augmentation

All images have been preprocessed and augmented prior to inclusion in the dataset. Users may apply additional transformations on top of these.

Preprocessing Applied

Step	Details
Auto-Orient	EXIF-based orientation correction applied
Resize	All images stretched to 224 × 224 pixels

Augmentation Applied (Training Set Only)

Augmentation	Setting
Outputs per example	2 (each training image yields 2 augmented variants)
Horizontal Flip	Enabled
Vertical Flip	Enabled
Rotation	Random angle in range [−15°, +15°]

✅ You are free to apply additional preprocessing or augmentation techniques based on your specific model requirements.

Usage

Loading with Hugging Face `datasets`

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/alzheimers-classification-dataset")

# Inspect a sample
print(dataset["train"][0])
# {'image': <PIL.Image>, 'label': 0}

# Class names
print(dataset["train"].features["label"].names)
# ['NonDemented', 'VeryMildDemented', 'MildDemented', 'ModerateDemented']

Loading from CSV files manually

import pandas as pd
from PIL import Image

train_df = pd.read_csv("CSV_datafiles/train.csv")
print(train_df.head())

# Load an image
img = Image.open(train_df.iloc[0]["filename"])
img.show()

Example: Fine-tuning with `transformers`

from datasets import load_dataset
from transformers import AutoFeatureExtractor, AutoModelForImageClassification, TrainingArguments, Trainer

dataset = load_dataset("YOUR_USERNAME/alzheimers-classification-dataset")

model_checkpoint = "google/vit-base-patch16-224"
extractor = AutoFeatureExtractor.from_pretrained(model_checkpoint)

def preprocess(batch):
    inputs = extractor(images=batch["image"], return_tensors="pt")
    inputs["labels"] = batch["label"]
    return inputs

processed = dataset.map(preprocess, batched=True, remove_columns=["image"])

model = AutoModelForImageClassification.from_pretrained(
    model_checkpoint,
    num_labels=4,
    id2label={0: "NonDemented", 1: "VeryMildDemented", 2: "MildDemented", 3: "ModerateDemented"},
    label2id={"NonDemented": 0, "VeryMildDemented": 1, "MildDemented": 2, "ModerateDemented": 3},
    ignore_mismatched_sizes=True,
)

training_args = TrainingArguments(
    output_dir="./alzheimers-vit",
    per_device_train_batch_size=32,
    num_train_epochs=10,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    metric_for_best_model="accuracy",
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=processed["train"],
    eval_dataset=processed["validation"],
)

trainer.train()

Considerations for Using the Data

Social Impact

Automated Alzheimer's staging tools have significant potential in clinical settings, particularly for:

Early diagnosis in resource-limited healthcare environments
Screening assistance to support radiologists and neurologists
Research acceleration in neurodegenerative disease studies

Responsible development of models trained on this data should involve clinical validation before any real-world deployment.

Discussion of Biases

Class imbalance: The ModerateDemented class is substantially underrepresented compared to NonDemented. Models trained without compensation (e.g., weighted loss, oversampling) may be biased toward majority classes.
Augmentation leakage risk: Since training augmentations are pre-applied, ensure no augmented variants of training images appear in the validation or test sets.
Demographic representation: The underlying OASIS data may not be representative of all ethnic, geographic, or socioeconomic populations, which could affect model generalizability.

Other Known Limitations

All images are resized to 224×224 via stretching (not padding), which may distort original spatial proportions.
This dataset is intended for research and educational purposes only and should not be used as the sole basis for clinical diagnosis.

Additional Information

Original Dataset Citation

If you use this dataset, please cite the original OASIS project:

@article{marcus2007open,
  title     = {Open Access Series of Imaging Studies (OASIS): Cross-Sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults},
  author    = {Marcus, Daniel S and Wang, Tracy H and Parker, Jamie and Csernansky, John G and Morris, John C and Buckner, Randy L},
  journal   = {Journal of Cognitive Neuroscience},
  volume    = {19},
  number    = {9},
  pages     = {1498--1507},
  year      = {2007},
  publisher = {MIT Press},
  doi       = {10.1162/jocn.2007.19.9.1498}
}

Kaggle Dataset Reference

kanaadlimaye. (2024). Alzheimer's Classification Dataset. Kaggle.
https://www.kaggle.com/datasets/kanaadlimaye/alzheimers-classification-dataset

OASIS Project

Website: https://www.oasis-brains.org/ Supported by: NIH grants P50 AG05681, P01 AG03991, R01 AG021910, P50 MH071616, U24 RR021382, R01 MH56584.

Contributions

Dataset curation, preprocessing, and augmentation by kanaadlimaye. Original MRI data provided by the OASIS project and its supporting institutions.

This dataset card was prepared following the Hugging Face Dataset Card guidelines.