Skip to content

index.rst

Latest commit

 

History

History
225 lines (159 loc) · 6.13 KB

index.rst

File metadata and controls

225 lines (159 loc) · 6.13 KB

Segmentation Modules

.. currentmodule:: physiomotion4d

AI-powered anatomical structure identification from medical images using state-of-the-art deep learning models.

Overview

PhysioMotion4D supports multiple segmentation approaches:

  • TotalSegmentator: Whole-body CT segmentation (100+ structures)
  • VISTA-3D: MONAI-based foundation model for medical imaging
  • VISTA-3D NIM: NVIDIA Inference Microservice version
  • Ensemble: Combine multiple methods for improved accuracy

All segmentation classes inherit from :class:`SegmentAnatomyBase` and provide consistent interfaces.

Quick Links

Segmentation Classes:

Choosing a Method

Method Speed Accuracy Best For
TotalSegmentator Fast (~30s) Good General purpose
VISTA-3D Medium (~60s) Excellent Cardiac imaging
VISTA-3D NIM Fast (cloud) Excellent Production
Ensemble Slow (~90s) Best Research/QC

Quick Start

Basic Segmentation

from physiomotion4d import SegmentChestTotalSegmentator

# Initialize segmentator
segmentator = SegmentChestTotalSegmentator(fast=True, verbose=True)

# Segment image
labelmap = segmentator.segment("ct_scan.nrrd")

# Extract specific structure
heart = segmentator.extract_structure(labelmap, "heart")

# Save results
segmentator.save_labelmap(labelmap, "output_labels.mha")

With VISTA-3D

from physiomotion4d import SegmentChestVista3D

# Initialize with GPU
segmentator = SegmentChestVista3D(
    device="cuda:0",
    use_auto_prompts=True,
    verbose=True
)

# Segment specific structures
labelmap = segmentator.segment(
    image_path="cardiac_ct.nrrd",
    structures=["heart_left_ventricle", "heart_myocardium"]
)

Ensemble Approach

from physiomotion4d import SegmentChestEnsemble

# Combine multiple methods
segmentator = SegmentChestEnsemble(
    methods=['totalsegmentator', 'vista3d'],
    fusion_strategy='voting',
    verbose=True
)

labelmap = segmentator.segment("ct_scan.nrrd")

Module Documentation

.. toctree::
   :maxdepth: 2

   base
   totalsegmentator
   vista3d
   vista3d_nim
   ensemble

Common Operations

Structure Extraction

Extract individual anatomical structures from segmentation results:

# Segment entire image
labelmap = segmentator.segment("ct.nrrd")

# Extract cardiac structures
lv = segmentator.extract_structure(labelmap, "heart_left_ventricle")
rv = segmentator.extract_structure(labelmap, "heart_right_ventricle")
myocardium = segmentator.extract_structure(labelmap, "heart_myocardium")

# Compute volumes
lv_volume = segmentator.compute_volume(lv)
print(f"LV volume: {lv_volume} mm³")

Batch Processing

Process multiple images efficiently:

from pathlib import Path

segmentator = SegmentChestTotalSegmentator(fast=True)

for image_file in Path("data").glob("*.nrrd"):
    print(f"Segmenting {image_file}...")

    labelmap = segmentator.segment(str(image_file))

    output_file = f"{image_file.stem}_labels.mha"
    segmentator.save_labelmap(labelmap, output_file)

Quality Control

Validate segmentation quality:

import numpy as np

def validate_segmentation(labelmap, ground_truth):
    """Compute Dice coefficient."""
    intersection = np.logical_and(labelmap, ground_truth).sum()
    dice = 2 * intersection / (labelmap.sum() + ground_truth.sum())
    return dice

# Validate results
dice_score = validate_segmentation(labelmap, reference_labelmap)
print(f"Dice score: {dice_score:.3f}")

Best Practices

Method Selection

  • TotalSegmentator: Use for general-purpose segmentation, fast iterations
  • VISTA-3D: Use for cardiac structures, when accuracy is critical
  • Ensemble: Use when maximum accuracy is needed, can tolerate longer processing

Parameter Tuning

  • Start with default parameters
  • Enable fast mode for quick prototyping
  • Use GPU (device="cuda:0") for VISTA-3D when available
  • Adjust post-processing based on image quality

Error Handling

try:
    labelmap = segmentator.segment(image_path)
except RuntimeError as e:
    print(f"Segmentation failed: {e}")
    # Fallback to alternative method
    segmentator_backup = SegmentChestTotalSegmentator(fast=True)
    labelmap = segmentator_backup.segment(image_path)

Performance Tips

  • Use GPU acceleration when available
  • Enable fast mode for development
  • Process time series in batch
  • Cache segmentation results for repeated use

See Also

Navigation

:doc:`../index` | :doc:`base` | :doc:`totalsegmentator`