Representation Learning in PET Scans Enhanced by Semantic and 3D Position Specific Characteristics.

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Representation learning is crucial for analyzing complex 3D medical images like FDG-PET.
• Existing methods face challenges due to computational costs and subtle image variations.
• Accurate classification of high-uptake regions is vital for diagnosing cancers such as Metastatic Melanoma.

Purpose of the Study:

• To develop a novel representation learning scheme for 3D FDG-PET images.
• To enhance the distinction between tumor and non-tumor regions.
• To improve diagnostic accuracy for Metastatic Melanoma.

Main Methods:

• Proposed a position-enhanced learning scheme incorporating semantic and position-based features.
• Introduced a Position Encoding Block (PEB) to generate informative representations.
• Evaluated the method on classifying high-uptake regions in FDG-PET scans for Metastatic Melanoma.

Main Results:

• Achieved a 10.50% increase in sensitivity and a 4.89% increase in F1-score compared to baseline methods.
• Demonstrated superior performance against state-of-the-art methods in classifying Metastatic Melanoma regions of interest.
• Validated the approach on both in-house and public whole-body FDG-PET datasets.

Conclusions:

• The proposed representation learning scheme effectively extracts generalizable features from high-dimensional medical data.
• This method significantly improves the classification of tumorous regions in FDG-PET scans for Metastatic Melanoma.
• The approach offers a promising advancement for clinical applications in cancer diagnostics.