A Comparison of Lung Nodule Segmentation Algorithms: Methods and Results from a Multi-institutional Study

Area of Science:

• Radiology
• Medical Imaging
• Computational Pathology

Background:

• Accurate tumor volume estimation and border segmentation in medical images are vital for cancer diagnosis, staging, and therapy response assessment.
• Computerized segmentation algorithms are increasingly used for lung nodule analysis in CT images.

Purpose of the Study:

• To assess the feasibility of a multi-institutional study evaluating the repeatability and reproducibility of lung nodule border segmentation and volume estimation bias of computerized algorithms.
• To compare the performance of different lung nodule segmentation algorithms using CT images.

Main Methods:

• Utilized a dataset of 52 tumors from 41 CT volumes from The Cancer Imaging Archive, including patient data and phantom nodules.
• Three academic institutions submitted results from three repeat runs for each nodule using their segmentation algorithms.
• Compared algorithm performance using spatial overlap measurements and volume estimation bias on both clinical and phantom datasets.

Main Results:

• Spatial overlap agreement was higher for repeat runs of the same algorithm compared to different algorithms (p < 0.05).
• Agreement was significantly higher on phantom datasets than on clinical datasets (p < 0.05).
• Algorithms showed significant differences in volume estimation bias for phantom nodules (p < 0.05), and high accuracy did not always correlate with high repeatability.

Conclusions:

• Multi-institutional assessment of segmentation algorithms is feasible.
• Significant variability exists between algorithms, particularly for heterogeneous nodules, necessitating the use of consistent software in longitudinal studies.
• Both accuracy and precision are critical when evaluating segmentation algorithms, and performance should be assessed on clinical data beyond phantoms.