Predicting distant metastases in soft-tissue sarcomas from PET-CT scans using constrained hierarchical multi-modality...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Soft-tissue sarcomas (STSs) management relies heavily on imaging modalities like Positron Emission Tomography-Computed Tomography (PET-CT).
• Distant metastases (DM) are the primary cause of mortality in STS patients, necessitating early detection for effective treatment.
• Current methods for DM detection in STS using PET-CT data can be improved for earlier and more accurate identification.

Purpose of the Study:

• To develop and evaluate a novel convolutional neural network (CNN) for the early detection of distant metastases (DM) in soft-tissue sarcoma (STS) patients.
• To leverage integrated functional (PET) and anatomical (CT) imaging features for enhanced DM detection accuracy.
• To create an automated system that does not require manual input, such as tumor delineation, for feature extraction.

Main Methods:

• Development of a constrained hierarchical multi-modality feature learning approach using a CNN.
• Integration of PET and CT imaging data within the CNN framework.
• Automated feature extraction without manual tumor delineation.

Main Results:

• The proposed CNN method achieved superior performance on a benchmark PET-CT dataset.
• The method demonstrated the highest accuracy (0.896) and Area Under the Curve (AUC) (0.903) compared to existing state-of-the-art techniques.
• Statistical analysis confirmed the significance of the results (unpaired student's t-test p-value < 0.05).

Conclusions:

• The developed CNN method is a promising tool for early DM detection in STS patients.
• This approach can assist clinicians in tumor quantification and the identification of imaging biomarkers for cancer treatment.
• The automated nature of the method enhances its potential for clinical application and support.