Efficient visual-search model observers for PET

Area of Science:

• Medical Imaging
• Radiology
• Computational Modeling

Background:

• Scanning model observers are established tools for predicting human performance in F-18 positron emission tomography (PET).
• Evaluating alternative model observers is crucial for improving prediction accuracy and efficiency in PET imaging analysis.

Purpose of the Study:

• To investigate the reliability and efficiency of a visual-search (VS) observer model compared to a channelized non-prewhitening (CNPW) scanning observer for predicting human performance in PET tumor detection.
• To assess whether VS observers can provide more stable and accurate predictions, especially under conditions of uncertainty.

Main Methods:

• Simulated 2D PET images with tumors in various organs and display formats were used.
• Localization receiver operating characteristic (LROC) studies quantified detectability for human observers and two model observers (CNPW and VS).
• VS observers utilized watershed and gradient-based search processes, evaluated under different background assumptions (BKE, BHA) and search areas (Task A, Task B).

Main Results:

• VS observers showed higher and more consistent Pearson correlations with human performance (Task A: 0.92-0.93, Task B: 0.92-0.93) compared to the CNPW observer (Task A: 0.77, Task B: 0.92) under the BKE assumption.
• The watershed-based VS observer processed 624 images in 2.0 minutes, while the CNPW observer processed images in 0.7 minutes, indicating comparable efficiency.

Conclusions:

• Computationally efficient VS models enhance the stability of statistical model observers for PET tumor detection tasks.
• VS models demonstrate improved concordance with human observers, suggesting their utility for more reliable performance predictions in PET imaging.