MULTI-TASK DEEP LEARNING AND UNCERTAINTY ESTIMATION FOR PET HEAD MOTION CORRECTION.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Head motion during brain PET imaging degrades image quality and introduces quantification errors.
• Previous Deep Learning Head Motion Correction (DL-HMC) showed promise but could be further improved.
• Accurate motion correction is crucial for reliable PET scan analysis.

Purpose of the Study:

• To develop and evaluate an upgraded multi-task deep learning model (mtDL-HMC) for enhanced head motion correction in brain PET.
• To integrate image appearance prediction into the deep learning framework for improved motion prediction.
• To assess the reliability and impact of discarding motion predictions with high uncertainty.

Main Methods:

• Developed a multi-task deep learning architecture (mtDL-HMC) incorporating image appearance prediction.
• Trained the mtDL-HMC model on data from 21 subjects.
• Evaluated performance on 5 testing subjects using quantitative and qualitative metrics.
• Employed Monte Carlo Dropout for assessing prediction uncertainty at inference.

Main Results:

• The mtDL-HMC model demonstrated superior motion prediction performance compared to the previous DL-HMC method.
• Both quantitative and qualitative assessments confirmed the enhanced accuracy of the mtDL-HMC.
• Discarding data with high motion prediction uncertainty did not compromise, and potentially improved, reconstructed image quality.

Conclusions:

• The multi-task deep learning approach significantly advances head motion correction in brain PET imaging.
• Integrating image appearance prediction and uncertainty assessment enhances the robustness and reliability of motion correction.
• This method offers a pathway to higher quality and more accurate quantitative PET analyses.