Automated detection and classification of tumor histotypes on dynamic PET imaging data through machine-learning...

Area of Science:

• Oncology
• Medical Imaging
• Machine Learning

Background:

• 2-deoxy-2-fluorine-(18F)fluoro-d-glucose Positron Emission Tomography/Computed Tomography (18F-FDG-PET/CT) is crucial for cancer diagnosis and staging.
• Histopathologic subtypes of lung cancer exhibit varying 18F-FDG uptake, leading to diagnostic limitations.
• Inflammatory processes and post-treatment changes can cause false-positive results in 18F-FDG-PET/CT scans.

Purpose of the Study:

• To develop a machine learning algorithm for automated classification of lung adenocarcinoma and other tumor types.
• To enhance the diagnostic accuracy and specificity of 18F-FDG-PET/CT in lung cancer.
• To improve the identification of metastatic lymph nodes in lung cancer patients.

Main Methods:

• A model-free, machine-learning based algorithm utilizing dynamic PET data (dPET) acquisitions.
• A trained Random Forest classifier analyzing spatial and temporal features of 18F-FDG uptake kinetics.
• Generation of probability maps for distinguishing adenocarcinoma from other lung histotypes and identifying metastatic lymph nodes.

Main Results:

• The algorithm achieved a probability of 0.943 ± 0.090 for detecting adenocarcinoma in a dPET dataset of 19 primary lung cancer patients.
• Demonstrated improved localization and discrimination of tumors.
• Showcased potential for accurately assessing tumor spread to the lymphatic system.

Conclusions:

• The proposed machine learning algorithm offers an automated and more accurate method for lung cancer diagnosis and staging.
• It enhances the specificity of 18F-FDG-PET/CT by differentiating tumor types and identifying metastatic lymph nodes.
• This approach provides a powerful tool for evaluating the extent of tumor metastasis in the lymphatic system.