Évaluation de méthodes automatiques de segmentation des volumes tumoraux en tomographie par émission de positons par comparaison avec des contours manuels réalisés par un groupe d’experts

Objective

Set up a framework for evaluating automatic segmentation methods of tumour volumes on PET images.

Patient and methods

This study was performed with PET images of 18 patients with non-Hodgkin's lymphoma. One target lesion per patient was pointed out. Each lesion was then three times manually delineated by five experts. Four automatic methods (the application of a threshold of 42% of the maximum SUV, the MIP-based method, the Daisne et al. method and the Nestle et al. method) were evaluated by comparison with the set of manual delineations.

Results

From the manual delineations, we have concluded to a moderate intra-operator variability and to a reduced interoperator reproducibility. From statistical tests performed on various quantitative criteria, there was no significant difference between the MIP-based method, the Daisne et al. method and the Nestle et al. one. The application of a threshold of 42% of the maximum SUV appears to be less efficient.

Conclusion

This work proposes a comparison and an evaluation protocol for segmentation methods. The generated data set will be distributed online for the community to simplify the evaluation of any new method of segmentation.