NeuroScan-AI

Inspiration

Alzheimer’s disease is one of the most challenging neurological disorders to diagnose early. While brain MRI scans contain valuable structural information, interpreting these scans requires significant expertise and time. During our research, we observed that many machine learning solutions focus heavily on achieving high accuracy, often overlooking interpretability and real-world clinical relevance.

This project was inspired by a simple but important question:

Can I design an AI system that not only detects Alzheimer’s disease from MRI scans but also explains its decisions in a way that clinicians can trust?

My goal was to move beyond black-box predictions and build a system that supports medical professionals rather than attempting to replace them.

What it does

I developed an end-to-end Convolutional Neural Network (CNN) pipeline that automatically classifies brain MRI scans into two categories:

Healthy or Alzheimer’s

To reflect realistic screening scenarios, I reformulated the original multi-class dataset into a binary classification task, focusing on the presence or absence of Alzheimer’s disease rather than fine-grained disease staging.

The system processes raw MRI data, extracts meaningful spatial features using CNNs, and produces predictions along with confidence scores and visual explanations using Grad-CAM.

How we built it

The project was built in the following stages:

Data Processing MRI scans stored in Parquet format were decoded from raw bytes, converted to grayscale, resized to a uniform resolution of 128 × 128 and normalized to ensure stable training.

Model Design I designed a compact CNN architecture consisting of convolutional layers for feature extraction, max-pooling layers for dimensionality reduction, and fully connected layers for classification. Dropout regularization was applied to reduce overfitting.

Training Strategy The model was trained using binary cross-entropy loss and the Adam optimizer. Early stopping and learning rate scheduling were employed to improve generalization. Class imbalance was addressed using class weighting to prevent bias toward the majority class.

Evaluation and Explainability Beyond standard metrics, I performed visual testing by inspecting individual predictions, including both correct and incorrect cases. Grad-CAM was applied to highlight brain regions that most influenced the model’s decisions, improving transparency and trust.

Challenges we ran into

Several challenges were encountered during development:

Data complexity: Handling MRI images stored in Parquet format required careful decoding and preprocessing.

Class imbalance: The dataset contained significantly fewer healthy samples, necessitating the use of class weighting.

Performance variability: Model accuracy varied across runs due to random initialization, requiring careful validation and consistency checks.

Interpretability: Ensuring that the model’s decisions were explainable and medically meaningful required additional design and evaluation effort.

Overcoming these challenges led to a more robust and credible system.

Accomplishments that we're proud of

Built a complete end-to-end medical imaging pipeline. I successfully developed a full workflow starting from raw MRI data stored in Parquet format to a trained and evaluated CNN model, handling decoding, preprocessing, training, evaluation, and visualization within a single reproducible system.

Reformulated a complex medical problem into a clinically realistic task By converting a multi-class Alzheimer’s dataset into a binary screening problem (Healthy vs Alzheimer’s), I aligned the model’s objective with real-world clinical screening needs rather than overfitting to artificial label granularity.

Achieved strong and consistent performance without overclaiming The model demonstrated high test accuracy (96%) across multiple runs while maintaining realistic expectations about performance variability, avoiding cherry-picked results and emphasizing consistency over a single best score.

Implemented explainable AI using Grad-CAM I integrated Grad-CAM visualizations to highlight the brain regions influencing model predictions, transforming the CNN from a black box into a transparent decision-support tool suitable for medical contexts.

What we learned

Throughout this project, I learned that accuracy alone is not sufficient, especially in medical AI applications. Important insights include:

Deep learning models can achieve high performance while still behaving unpredictably without proper evaluation.

Visual error analysis provides valuable insight into model behavior that numerical metrics cannot capture.

Explainability methods such as Grad-CAM are essential for building trust in healthcare-oriented AI systems.

Slight performance variation across runs is normal due to stochastic training and should be acknowledged rather than hidden.

These lessons reinforced the importance of responsible AI development in sensitive domains like healthcare.

What's next for NeuroScan-AI

This project demonstrates how explainable deep learning models can assist in the early screening of Alzheimer’s disease using brain MRI scans. While the system is not intended for clinical diagnosis, it provides a strong foundation for decision-support tools in medical imaging.

Future work could include:

External validation on additional datasets

Extension to multi-stage disease classification

Integration into clinical imaging workflows