A statistical method for lung tumor segmentation uncertainty in PET images based on user inference.

Area of Science:

• Medical Imaging
• Radiology
• Computational Biology

Background:

• Positron Emission Tomography (PET) is crucial for lung tumor diagnosis and treatment.
• Standardized criteria for delineating tumor boundaries in PET are lacking due to image quality and tumor variability.
• Segmentation uncertainty remains a challenge in accurate lung tumor characterization.

Purpose of the Study:

• To develop a statistical method for addressing segmentation uncertainty in lung tumor delineation on PET images.
• To integrate user inference into a quantitative framework for improved boundary definition.
• To enhance the reliability of PET imaging in lung cancer management.

Main Methods:

• A novel statistical approach was developed to define an uncertainty segmentation band using a Random Walks (RW) algorithm.
• Principle Component Analysis (PCA) was employed to formulate a statistical model for labeling the uncertainty band based on user-extracted features.
• The method was validated on 10 lung PET-CT phantom studies and 16 clinical PET studies.

Main Results:

• The proposed method demonstrated robust performance in delineating lung tumor boundaries.
• An average Dice Similarity Coefficient (DSC) of 0.878 ± 0.078 was achieved on phantom studies.
• An average DSC of 0.835 ± 0.039 was obtained on clinical studies, indicating high spatial overlap accuracy.

Conclusions:

• The statistical solution effectively reduces segmentation uncertainty in lung tumor delineation on PET images.
• The integration of user inference provides a valuable tool for improving the accuracy of PET-based tumor boundary definition.
• This method holds promise for enhancing diagnostic and therapeutic precision in lung cancer patients.