Sequential deep learning image enhancement models improve diagnostic confidence, lesion detectability, and image...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiochemistry

Background:

• Deep learning (DL) algorithms, DL-Enhancement (DLE) and DL-based time-of-flight (DLT), were investigated for sequential application.
• DLE enhances ordered-subset-expectation-maximisation (OSEM) images towards block-sequential-regularised-expectation-maximisation (BSREM) quality.
• DLT improves BSREM images reconstructed without time-of-flight (ToF).

Purpose of the Study:

• To evaluate the combined benefits of sequentially applying DLE and DLT in PET-CT imaging.
• To assess the impact on image quality, lesion detectability, and diagnostic confidence.
• To determine the effect on image reconstruction time.

Main Methods:

• Forty FDG PET-CT scans were acquired on GE Healthcare Discovery 710 (D710) and Discovery MI (DMI) scanners.
• Five reconstruction combinations were tested: ToF-BSREM, ToF-OSEM + DLE, OSEM + DLE + DLT, ToF-OSEM + DLE + DLT, and ToF-BSREM + DLT.
• Image noise was measured using spherical VOIs in the lung and liver. Clinical readers assessed lesion detectability, diagnostic confidence, and image quality using a Likert scale.

Main Results:

• The sequential application of DLE + DLT reduced noise and improved lesion detectability, diagnostic confidence, and reconstruction time.
• ToF-OSEM + DLE + DLT reconstructions showed increased lesion SUVmax by 28% (D710) and 11% (DMI).
• This reconstruction method achieved the highest scores for lesion detectability and diagnostic confidence in clinical readings for D710 data.

Conclusions:

• Combining DLE and DLT enhances diagnostic confidence and lesion detectability compared to standard ToF-BSREM.
• The rapid DL inferencing in DLE + DLT significantly reduced overall reconstruction time.
• This approach holds potential for improving total body PET imaging applications.