Medical Imaging Research · Notebook-based study

Alzheimer’s Disease MRI Research

A Google Colab MRI classification study that keeps image inspection, preprocessing, staged labels, CNN training, and comparison visible in the notebook record.

Problem

Medical imaging experiments can become difficult to judge when the final score is separated from the images, preprocessing choices, label structure, and evaluation path.

Approach

The notebook keeps the study sequence visible: MRI inspection, image resizing, augmentation, four-class labeling, CNN training, and ensemble-oriented comparison.

Output

A repository-backed Colab notebook and supporting image outputs that show the experiment as a reproducible sequence rather than a polished clinical claim.

Role: I built the data preparation, augmentation, training, and evaluation flow and preserved the notebook as the visible record of the experiment.

Study frame

A medical imaging study preserved as an experimental record

The work stays close to its actual experimental form: a notebook-driven MRI classification study in Google Colab that moves from image inspection and preprocessing to model training and comparison.

The notebook is treated as the primary record of the work because it preserves the order of decisions, not only the final result.

Data view

Representative MRI slices keep the data visible

The notebook includes direct visual inspection of MRI inputs. Keeping examples like these in the workflow matters because it ties preprocessing and model design back to the actual structure of the data rather than treating the dataset as an abstraction.

MRI sample montage extracted from the notebook output in the repository.

Problem framing

The classification problem is staged, not simply binary

The classification target is more demanding than a simple healthy-versus-disease split.

The study uses four classes: NonDemented, VeryMildDemented, MildDemented, and ModerateDemented. That framing matters because the model is not only asked to separate apparent disease from no disease. It is asked to distinguish degrees of progression that can be visually close and unevenly represented.

The notebook reflects that framing at each stage. Data preparation, augmentation, and evaluation sit inside a multi-class setup, and the later ensemble work only makes sense because a visible baseline has already been established. The value of the project is not a claim of clinical readiness. It is the preserved experimental sequence: what data was used, how images were prepared, how the model was trained, and how later comparison entered the workflow.

Pipeline

Key experimental choices

4Classification labelsNonDemented, VeryMildDemented, MildDemented, and ModerateDemented.

256×256Image sizingThe notebook resizes images before model training.

80/10/10Data splitTraining, validation, and test partitions described in the repository README.

CNN + ensembleModeling approachA baseline convolutional model is extended with simple ensemble comparison.

Why the format matters

Notebook-first work retains useful forms of visibility

In this study, inspectability is part of the value rather than a sign of incompleteness.

Data

Images remain visible

Representative MRI slices stay inside the experimental record, which keeps preprocessing and modeling grounded in the data itself.

Method

Pipeline decisions stay local

Loading paths, resizing, augmentation, training, and evaluation appear in one continuous environment instead of being scattered across scripts.

Interpretation

Comparisons can be read in sequence

Later ensemble or comparison logic is easier to judge when it appears as an extension of earlier notebook steps rather than as an isolated result.

Preprocessing

Single-slice inspection helps keep the pipeline concrete

Notebook-based work is often criticized for being less polished than packaged code, but it has one important advantage: the data remains visible. Images like this make preprocessing decisions easier to inspect and discuss.

Single MRI slice extracted from the repository notebook output.

Method

The notebook is the record of the study

Its value lies in preserving sequence and decision-making, not in pretending the work happened somewhere else.

In this project, the notebook is not treated as an intermediate artifact waiting to become a cleaner library. It is the most faithful representation of the work. It shows where the dataset is loaded, how the images are augmented, how the convolutional model is trained, and how later comparison logic is layered in.

That visibility is important in medical and scientific contexts. The useful question is not whether a result looks good in isolation, but whether the sequence of choices behind it can be examined with enough detail to judge the work responsibly.

Experimental flow

How the study progresses

The notebook format makes each stage legible in order rather than compressing everything into a final result.

Load and inspect MRI inputs

The first step is not training but looking at the images closely enough to understand what the data actually contains.

Prepare and augment the dataset

Resizing, partitioning, and augmentation decisions are made in the open, with the implications visible alongside the code.

Train the baseline CNN

The core convolutional model establishes the initial performance frame for the classification task.

Extend into comparison and ensemble logic

Later evaluation becomes meaningful because it grows directly out of the earlier pipeline rather than replacing it.

Evaluation context

The project is best understood as a visible sequence of experiments

The later portions of the notebook move beyond a single model result and into comparison. That transition is where the study becomes more interesting, because it reveals how performance was interrogated rather than simply reported.

Additional MRI output extracted from the notebook record.

Scope

What this study demonstrates

The contribution is methodological transparency rather than a claim of production or clinical readiness.

This work demonstrates a practical medical imaging workflow built under notebook-first conditions: dataset preparation, convolutional modeling, staged classification, and comparison-oriented evaluation in a single reproducible file. It is not presented as a clinical system or a polished deployment target. Its value lies in the transparency of the experimental process and in the way it captures model development in a form that can still be reviewed, rerun, and extended.

Repository source

README appendix

The curated project page above is the main reading path. The imported README is kept as source context for readers who want to compare the portfolio narrative with the repository record.

Open imported README

Alzheimer's Disease Detection with Ensemble Learning

This repository contains a Google Colab notebook for classifying Alzheimer's disease progression from brain MRI images using convolutional neural networks and ensemble learning experiments.

The project focuses on four diagnostic categories derived from Clinical Dementia Rating style staging:

NonDemented
VeryMildDemented
MildDemented
ModerateDemented

Project Overview

The notebook builds an end-to-end image classification workflow:

load MRI image data from Google Drive
create a TensorFlow dataset pipeline
split the data into training, validation, and test sets
apply resizing, rescaling, and augmentation
train a CNN classifier
evaluate predictions on held-out samples
compare single-model results with ensemble variants built from saved models

The notebook is exploratory and notebook-first. It is best read as an experimentation workflow rather than a packaged training library.

Notebook Contents

The main notebook is AZclassification5(wide).ipynb.

It includes:

background notes on why early Alzheimer's detection matters
dataset framing for the four MRI classes
TensorFlow/Keras preprocessing and augmentation
a baseline sequential CNN model
sample prediction visualizations
loading previously saved models from Google Drive
simple averaging ensembles across multiple trained models
an initial weighted-ensemble section

Model Pipeline

Baseline CNN

The baseline model in the notebook uses:

input resizing to 256 x 256
pixel rescaling to [0, 1]
data augmentation with flips, rotations, zoom, width shifts, and height shifts
stacked Conv2D + MaxPooling2D blocks
dropout before the dense classifier head
a final softmax layer over 4 classes

Dataset Split

The TensorFlow pipeline partitions the dataset into:

80% training
10% validation
10% test

Ensemble Experiments

Later sections of the notebook load additional saved models from Google Drive and test:

a 2-model averaging ensemble
a 3-model averaging ensemble
a draft weighted-ensemble experiment

Tech Stack

Python 3
TensorFlow / Keras
NumPy
OpenCV
Matplotlib
PIL
Google Colab

The notebook also installs and references:

opencv-python
keras_sequential_ascii
visualkeras

How To Run

Option 1: Run in Google Colab

Open the notebook with the badge above, then:

mount Google Drive
upload or link your MRI dataset in Drive
update the hardcoded dataset path in the notebook if needed
run the cells in order

Option 2: Run Locally

You can also run the notebook locally in Jupyter, but you will need to replace the Colab-specific parts:

remove or adapt google.colab.drive
update the dataset path from the current Google Drive location
ensure TensorFlow and the imaging dependencies are installed locally

Example package install:

pip install tensorflow opencv-python matplotlib pillow numpy visualkeras keras-sequential-ascii

Data Assumptions

The committed notebook expects a dataset path like:

data_dir = r"/content/drive/MyDrive/datasetzip/dataset/photos"

The dataset itself is not included in this repository, so you will need to provide your own MRI image folder structure before running the notebook.

Repository Structure

.
├── AZclassification5(wide).ipynb
└── README.md