wenmin-wu/cv-3d-patch-sliding-window-inference

3D Patch Sliding-Window Inference

Overview

Large 3D volumes (CT, cryo-ET, MRI) rarely fit in GPU memory whole. Sliding-window inference tiles the volume into overlapping patches, runs the model on each, then blends overlapping regions via averaging or Gaussian weighting. Eliminates boundary artifacts while keeping memory constant.

Quick Start

from monai.inferers import SlidingWindowInferer

inferer = SlidingWindowInferer(
    roi_size=(96, 96, 96),
    sw_batch_size=4,
    overlap=0.5,
    mode="gaussian",  # Gaussian weighting reduces edge artifacts
)
output = inferer(inputs=volume_tensor, network=model)

Workflow

1. Choose roi_size matching the model's training patch size
2. Set overlap (0.25–0.5) — higher overlap = smoother but slower
3. Select blending mode: "gaussian" weights center pixels more; "constant" averages uniformly
4. Run inferer — it handles tiling, batching, and stitching automatically
5. Post-process the full-resolution prediction volume

Key Decisions

• Overlap ratio: 0.5 is standard; 0.25 acceptable if speed-constrained
• Blending mode: Gaussian preferred — reduces checkerboard artifacts at patch boundaries
• sw_batch_size: Max patches per forward pass; tune to available VRAM
• Without MONAI: Implement manually with np.lib.stride_tricks + weighted accumulation buffer

References

• Baseline UNet train + submit
• CZII YOLO11+Unet3D-Monai LB.707
• 3d-unet using 2d image encoder