NeuroSpineAI: SD vs. MET MRI Classification

Inspiration: Spondylodiscitis (SD) and spinal metastases (MET) often appear radiologically similar in early stages, yet require completely different treatments. A delayed or incorrect diagnosis can lead to irreversible neurological damage. This project was personally inspired by a close friend who developed spinal infection complications that were not diagnosed early. After reviewing clinical literature and open-source datasets, we realized there is a real need for computational tools that support radiologists in distinguishing infectious from malignant spinal lesions.

We asked a simple question: Can machine learning assist in improving early differentiation between SD and MET using quantitative MRI features?

What It Does: NeuroSpineAI is a neural-network-based classifier that distinguishes Spondylodiscitis (SD) from Spinal Metastases (MET) using 31 radiomic features extracted from MRI lesions. The model was trained and evaluated on: Training/Test Set: 80 lesions (40 SD, 40 MET) Validation Set: 91 lesions (55 SD, 36 MET) The goal is to maximize diagnostic accuracy while maintaining generalization and preventing overfitting.

How I Built It: Model Architecture I implemented two feedforward neural network architectures: Model A 1 hidden layer 6 neurons tanh activation Model B 2 hidden layers 7 neurons (Layer 1) 3 neurons (Layer 2) tanh activation

The output layer used sigmoid activation for binary classification: 𝑦 = 𝜎(𝑊𝑥+𝑏)y=σ(Wx+b) The loss function used was binary cross-entropy: 𝐽=−1𝑚∑𝑖=1𝑚[𝑦(𝑖) log(𝑦(𝑖))+(1−𝑦(𝑖))log⁡(1−𝑦(𝑖))] J=− m1 i=1∑m[y(i)log(y(i))+(1−y(i))log(1−y(i))] Training Approaches: I compared three gradient-based learning modes: Batch Gradient Descent Stochastic Gradient Descent Mini-Batch Gradient Descent (b = 16) Hyperparameters tested: Iterations: 5,000 – 50,000 Learning rate: 0.001 – 0.01 Mini-batch size: 16

Results: Learning Mode Performance Batch Mode: Stable convergence; best at learning rate 0.0075 with 50,000 iterations Stochastic Mode: High variance; lower overall performance Mini-Batch Mode: Best balance of convergence speed and generalization Architecture Comparison Model A (1 Hidden Layer): Good baseline but limited in capturing complex patterns Model B (2 Hidden Layers): Higher accuracy and improved feature representation

The best-performing configuration was: 2 hidden layers + Mini-Batch (b=16) + learning rate 0.0075

Challenges I Faced: Small dataset size increased risk of overfitting Hyperparameter tuning required careful experimentation Feature scaling significantly affected convergence behavior Stochastic training introduced instability I learned how sensitive neural networks are to initialization, scaling, and training mode selection.

What We Learned: Mini-batch training provides the best tradeoff between stability and speed Deeper networks (even modestly deeper ones) improve nonlinear feature extraction Proper normalization is critical for medical datasets Model architecture matters as much as hyperparameters

Next Steps: To improve performance and clinical applicability: Replace normalization with standardization (zero mean, unit variance) Apply SelectKBest for feature selection Integrate Batch Normalization and Dropout Use Adam optimizer for adaptive learning rates Compare with XGBoost for better interpretability and feature importance ranking

Built With

• custom
• implementation
• jupyter
• matplotlib
• network
• neural
• notebook
• numpy
• pandas
• python
• scikit-learn