CANDI

CANDI facilitates collaborative annotation of radiographs and randomized evaluation of computer-aided diagnosis (CAD) impact on radiologist interpretation.

Key Features:

• Dual-component system: Separate annotation and evaluation components for collecting training data and conducting randomized assessments of CAD influence.
• Annotation data types: Collection of classification labels, segmentation masks, and image captions for radiographs.
• Randomized CAD availability: Randomizes the presence of CAD assistance during evaluations to measure unbiased effects on interpretation.
• Support for deep learning: Provides diverse labeled radiograph data to train and refine deep learning and machine learning algorithms.
• R package implementation: Distributed as an R package for computational integration with analysis workflows.
• Human–AI collaboration framework: Explicitly designed to study interactions between radiologists and CAD systems.

Scientific Applications:

• Training-data generation for deep learning: Produces classification, segmentation, and caption labels for supervised deep learning on radiographs.
• Randomized evaluation of CAD: Enables controlled trials to assess how CAD affects radiologist diagnostic accuracy and decision-making.
• Human–AI interaction studies: Supports investigations into collaborative decision-making between radiologists and algorithmic assistance.
• Assessment of CAD impact: Facilitates studies determining whether CAD enhances or hinders clinical interpretation of medical images.

Methodology:

Implemented as an R package; an annotation component collects classification labels, segmentation masks, and image captions, and an evaluation component conducts controlled clinical trials by randomizing CAD availability to assess effects on radiologist interpretation and to provide training data for machine learning and deep learning algorithms.