Operating characteristics predicted by models for diagnostic tasks involving lesion localization

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• The Swensson observer model unified ROC, LROC, FROC, and AFROC curves.
• Recent models include initial detection and candidate analysis (IDCA) for CAD data and a search model for human observer data.

Purpose of the Study:

• Derive IDCA and search model expressions for operating characteristics.
• Compare IDCA and search model predictions to the Swensson model.
• Evaluate model performance on mammography CAD datasets.

Main Methods:

• Derived mathematical expressions for operating characteristics based on IDCA and search models.
• Fitted four mammography CAD datasets using the Swensson, IDCA, and search models.
• Compared model fits, focusing on high- and low-confidence regions.

Main Results:

• All models fit high-confidence data well for three out of four datasets.
• IDCA and search models showed better fits in low-confidence regions for FROC curves.
• Swensson model predictions significantly diverged from data in low-confidence regions for FROC curves.
• A unique characteristic of IDCA and search models is the non-continuous movement of the operating point, observed in CAD data.

Conclusions:

• IDCA and search models offer improved performance for analyzing mammography CAD data, particularly in low-confidence regions.
• The Swensson model's limitations in low-confidence regions are explained by its continuous operating point prediction.
• Further research may refine these models for enhanced diagnostic accuracy in medical imaging.