Evaluation of PET volume segmentation methods: comparisons with expert manual delineations

Area of Science:

• Nuclear Medicine
• Medical Imaging Analysis
• Oncology

Background:

• [¹⁸F]-Fluorodeoxyglucose Positron Emission Tomography (FDG-PET) is crucial in cancer management.
• Accurate tumor segmentation in FDG-PET images is vital for effective treatment planning.
• A standardized evaluation tool is needed due to the proliferation of segmentation methods.

Purpose of the Study:

• To evaluate and compare the performance of five different FDG-PET image segmentation methods.
• To establish a reliable framework for assessing new segmentation techniques.
• To analyze intraoperator and interoperator variability in manual delineations.

Main Methods:

• Five segmentation methods were tested: Maximum Intensity Projection (MIP)-based, Fuzzy C-means, Daisne, Nestle, and a 42% threshold-based approach.
• Methods were compared against manual delineations by a panel of expert radiologists on non-Hodgkin's lymphoma lesions.
• Similarity measures, including the binary Jaccard index, were used for quantitative analysis.

Main Results:

• The MIP-based method demonstrated the highest similarity to expert manual delineations (Jaccard index: 0.45 ± 0.15).
• Fuzzy C-means yielded slightly lower performance, while the 42% threshold method performed the furthest from manual segmentation (Jaccard index: 0.38 ± 0.16).
• Significant intraoperator and interoperator variability were observed in the manual delineation process.

Conclusions:

• A simple assessment framework using expert manual delineations is proposed for evaluating segmentation methods.
• Utilizing multiple expert delineations effectively accounts for interobserver variability.
• The study's dataset will be shared online to facilitate the evaluation of novel segmentation techniques.