Deep learning techniques in PET/CT imaging: A comprehensive review from sinogram to image space.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Hybrid Positron Emission Tomography/Computed Tomography (PET/CT) imaging offers comprehensive data for oncology, neurology, and cardiology.
• Manual interpretation of PET/CT scans is time-consuming and requires specialized knowledge.
• Existing reviews on AI in medical imaging often lack focus on the specific application of deep learning (DL) to PET/CT.

Purpose of the Study:

• To comprehensively review studies applying deep learning (DL) to Positron Emission Tomography/Computed Tomography (PET/CT) imaging.
• To identify effective DL approaches, pre-processing techniques, and limitations in the field.
• To analyze the characteristics of DL applications in PET/CT image analysis.

Main Methods:

• Systematic review of 99 studies published between 2017 and 2022 focusing on DL in PET/CT.
• Analysis of pre-processing algorithms and deep learning models used in the selected studies.
• Identification of common and specific pre-processing techniques and their effectiveness.

Main Results:

• Deep learning demonstrates significant potential in PET/CT for lesion detection, tumor segmentation, and disease classification.
• Effective DL models and pre-processing techniques for PET/CT analysis were identified.
• Key limitations include scarcity of annotated datasets and challenges in model explainability and uncertainty quantification.

Conclusions:

• Deep learning significantly improves accuracy and efficiency in PET/CT image interpretation.
• Emerging DL models (e.g., attention-based, transformers) and radiomics show promise for advancing PET/CT applications.
• Continued research is essential to overcome limitations and expand the use of DL in PET/CT imaging.