NA-MIC Project Weeks

Back to Projects List

Slicer Workflows

Key Investigators

Description

Creating streamlined workflows in Slicer

Topics

Add ideas for discussion here!

Discussion topics:

• Best practices
• What these extensions have in common
• How to create the infrastructure to support the annotation workflow
• Tasks that a workflow should contain: curation, annotation, review, comparing multiple annotations, classification
• Involvement and use of other software like OHIF

CART Base Classes: Generic Iterator Framework for 3D Slicer

Overview

CART (Collaborative Annotation and Review Tool) provides a set of abstract base classes for creating streamlined annotation workflows in 3D Slicer. The framework enables efficient iteration through medical imaging cohorts with customizable tasks and flexible data loading strategies.

Contributors

• Ivan Johnson-Eversoll (University of Iowa, USA)
• Kalum Ost (Montreal Polytechnic, Canada)
• Kuan Yi (Montreal Polytechnic, Canada)

With inspiration from SlicerCART, mpReview, and CaseIterator

Design Philosophy

Minimal Requirements, Maximum Flexibility

The framework enforces only one requirement:

• A CSV file with a uid column for unique case identification

We chose CSV for its simplicity and universal usage - it’s human-readable, version-controllable, and supported by every data processing tool.

Pluggable Data Loading Architecture

The key innovation is the DataUnit abstraction layer that decouples data sources from task logic:

CSV (uid + metadata) → DataUnit → Task Logic
     ↓                    ↓           ↓
  Universal            Pluggable   Reusable
  Interface            Loaders     Tasks

This design enables multiple data loading strategies:

Local File Support

class VolumeOnlyDataUnit(DataUnitBase):
    # Loads .nrrd files from local filesystem
    # Supports relative paths with configurable base directory

DICOM Integration (Future)

class DICOMDataUnit(DataUnitBase):
    # Query DICOM database by SeriesInstanceUID
    # Automatic series type detection and loading
    # Built-in de-identification handling

Cloud/Remote Loading (Future)

class CloudDataUnit(DataUnitBase):
    # Download from S3, Google Cloud, or IDC
    # Automatic caching and prefetching
    # Authentication handling

Key Design Benefits

1. Task-Agnostic Data Loading

Tasks don’t need to know whether data comes from local files, DICOM databases, or cloud storage. They simply request resources through the standard get_resource(key) interface.

2. Data Source Migration

Switch from local NIfTI files to a DICOM database by changing only the DataUnit type - no task code changes required.

3. Parallel Development

Data engineers can optimize loading strategies while UI developers focus on annotation workflows, all working against the same abstract interface.

4. Custom Hanging Protocols

Tasks can define sophisticated view layouts and multi-volume displays without coupling to specific data formats.

Architecture Overview

Core Components

DataUnitBase (Abstract)

• Purpose: Abstract interface between CSV metadata and loaded Slicer data
• Responsibility: Validate data, load resources, manage MRML scene integration
• Extensibility: Subclass for different data sources (files, DICOM, cloud, etc.)

DataManager

• Purpose: CSV cohort management and efficient case navigation
• Features: Sliding window traversal, wraparound navigation, progress tracking
• Future: Multi-scene prefetching for seamless user experience

TaskBaseClass (Abstract)

• Purpose: Generic framework for annotation/review tasks
• Features: GUI integration, automatic data binding, save/resume capability
• Customization: Define custom hanging protocols and user interfaces

Prefetching and Multi-Scene Support

The framework is designed to support prefetching multiple scenes for improved performance:

Current Implementation

• Single scene with lazy loading
• Automatic scene clearing on navigation

Future Enhancement

• Background loading of next/previous cases
• Intelligent memory management
• Seamless transitions without loading delays

Custom Hanging Protocols

Tasks can implement sophisticated display logic:

class MultiContrastTask(TaskBaseClass):
    def setup(self, data_unit):
        # Custom layout: T1w background, T2w foreground, FLAIR in separate view
        volumes = [data_unit.get_resource(key) for key in ['T1w', 'T2w', 'FLAIR']]
        self.layoutLogic.viewerPerVolume(volumes, background=volumes[0])

CSV Format Specification

The only requirement is a uid column. All other columns are interpreted as resource identifiers. We support saving to the Input CSV or a separate output CSV.:

uid,T1w,T2w,segmentation,notes
patient_001,/data/001/t1.nrrd,/data/001/t2.nrrd,,needs_review
patient_002,/data/002/t1.nrrd,/data/002/t2.nrrd,/segs/002.nrrd,complete

Getting Started

1. Create Your DataUnit

class MyDataUnit(DataUnitBase):
    def _validate(self):
        # Ensure your data meets requirements
        
    def _initialize_resources(self):
        # Load data into Slicer scene
        
    def to_dict(self):
        # Export for saving back to CSV

2. Create Your Task

class MyTask(TaskBaseClass):
    def buildGUI(self, container):
        # Build your annotation interface
        
    def setup(self, data_unit):
        # Configure views and load data
        
    def save(self):
        # Save annotations/results

3. Prepare Your CSV

• Include uid column with unique identifiers
• Add columns for each resource (images, segmentations, etc.)
• Use paths appropriate for your chosen DataUnit implementation

Step 1

Get Started by simply selecting your CSV file in the module interface. And Select your “User” or add a new one. You can Then navigate through your data using the “Next” and “Previous” buttons.

Step 2

Select your task from the dropdown menu.

Step 3

And complete the “Action” you want to perform. And you can still use the same “Next” and “Previous” buttons to navigate through your data.