MONAI 3D U-Net Brain Segmentation

Deep Learning-Based Automated Brain Extraction from Post-Operative CT Scans

Architecture & Implementation

Network Structure

5 hierarchical levels
Channels: 32 → 64 → 128 → 256 → 512
Encoder-decoder with skip connections

Framework

MONAI 1.4.0 toolkit
PyTorch backend
Medical imaging specialized

Regularization

Dropout: 0.5 rate
Instance normalization
LeakyReLU (slope=0.01)

Input/Output

Single-channel 3D CT volume
Binary brain mask output
HU normalization [-1000,1000] → [0,1]

Data Preprocessing: I_normalized = clip((I_raw - (-1000))/2000, 0, 1) Binary Output: M_binary = I(σ(M_raw) > 0.5)

Performance Results

93.6% Final Model
Dice Score
90.8% External
Validation
40.1% Performance
Improvement
96.9% Variance
Reduction

Training Strategy

Component Configuration Purpose
Loss Function 0.8×Dice + 0.2×BCE Handles class imbalance
Optimizer AdamW: LR 1×10⁻⁴→1×10⁻⁵ Weight decay: 1×10⁻⁵
Data Augmentation Flipping (60%), Rotation (60%)
Gaussian noise (30%)		 Strategic pipeline
Validation Leave-One-Out Cross-validation 8-fold patient-wise split
Ground Truth 7 mask variants per patient Averaged for robustness

Clinical Impact