> loading citations module...

> parsing research references...

> indexing model architectures...

> compiling performance metrics...

> rendering interface ████████ OK

Dataset

Parkinson's Drawings — sourced from Kaggle. Contains spiral and wave drawings from healthy individuals and patients diagnosed with Parkinson's disease, captured via digitizing tablets with high-resolution coordinate tracking.

Two classes: healthy & parkinson
Two test modalities: Spiral drawings & Wave drawings

Model Architecture

Backbone

MobileNetV2 pre-trained on ImageNet. Feature extraction layers are frozen; only the custom fully-connected classification head is fine-tuned on the Parkinson's drawing dataset.

Classification Head

Global Average Pooling → Dense(128, ReLU) → Dropout(0.3) → Dense(2, Softmax)

Ensemble Strategy

Weighted ensemble: 40% Spiral / 60% Wave — weights derived from validation AUC performance.

Performance Metrics

Spiral CNN

Accuracy86.7%
AUC-ROC0.9511

Wave CNN

Accuracy91.3%
AUC-ROC0.9627

Tech Stack

Backend

FastAPI

High-performance Python web framework

TensorFlow 2.10

Deep learning inference engine

Keras

Model definition & training API

Frontend

Next.js 14

React framework with App Router

TypeScript

Type-safe JavaScript

Tailwind CSS

Utility-first styling

React / Canvas API

Interactive drawing capture

Research Papers

Disclaimer: NeuroSketch is for educational and screening purposes only. It is not a medical diagnostic device. Always consult a qualified healthcare professional for clinical evaluation.