A Bayesian approach to tissue-fraction estimation for oncological PET segmentation

Area of Science:

• Medical Imaging
• Oncology
• Artificial Intelligence

Background:

• Tumor segmentation in oncological PET is hindered by partial-volume effects (PVEs) and tissue-fraction effects (TFEs).
• Conventional methods struggle to model TFEs, where voxels contain mixed tissue classes.
• Accurate tumor segmentation is crucial for effective cancer treatment planning and monitoring.

Purpose of the Study:

• To develop and evaluate a novel Bayesian approach for tissue-fraction estimation in oncological PET segmentation.
• To address the limitations of conventional methods in handling PVEs and TFEs.
• To improve the accuracy and reliability of tumor segmentation in PET imaging.

Main Methods:

• A Bayesian approach was employed for tissue-fraction estimation, implemented via a deep-learning technique.
• The method estimates the posterior mean of tumor volume fraction within each voxel.
• Evaluated using 2D simulations and clinical data from stage IIB/III non-small cell lung cancer patients (ACRIN 6668/RTOG 0235).

Main Results:

• The proposed Bayesian method significantly outperformed conventional and U-net-based segmentation techniques in simulations and clinical data.
• Achieved a Dice Similarity Coefficient (DSC) of 0.82 (95% CI: 0.78, 0.86) on clinical patient images.
• Demonstrated high accuracy for small tumors, with a DSC of 0.77 for the smallest segmented cross-section (1.30 cm²).
• Showed robustness to PVEs and reliable performance across different scanner configurations.

Conclusions:

• The proposed Bayesian deep learning method effectively segments tumors in PET images, overcoming PVEs and TFEs.
• This approach offers a significant advancement in oncological PET image analysis.
• The method provides accurate and reliable tumor segmentation, particularly beneficial for small or challenging cases.