A new segmentation approach for F-18-fluoromisonidazole positron emission tomography data based on Ant Colony Optimization: Considering Reproducibility

Ziel/Aim:

Hypoxia imaging using F-18-fluoromisonidazole (FMISO) positron emission tomography (PET) is of increasing interest in the field of radiation oncology. But for analysis of FMISO PET data a reliable and accurate delineation technique of hypoxic subvolumes is still needed. Due to the inability of threshold based segmentation approaches to deliver reliable results when applied to data sets with small, inhomogeneous or non spherical target volumes, more complex algorithms may be preferred (1). We propose an Ant-Colony-Optimization (ACO) approach for segmentation of FMISO PET data sets. This investigation was performed to validate the reproducibility of the algorithm processing patient data sets.

Methodik/Methods:

Our analysis included 28 patients from an ongoing prospective study on head and neck cancers. FMISO PET images were acquired 4 hours p.i. of ~266 MBq. Patients were investigated by FMISO PET before radiochemotherapy (RCT) and the resulting data sets were further processed by the proposed ACO approach. Virtual ants were operating autonomously in the PET volume searching for regions with signal intensity above average. When an ant is located in such a region, it emits pheromone, attracting more ants to go to the marked region. More ants emit more pheromone and in that way the pheromone field shows the target objects with higher contrast than the original data set. Afterwards the pheromone fields were segmented into positive and negative regions using a histogram based threshold algorithm. Each data set was processed for 3 times and the resulting delineations were compared pair wise using the Jaccard-Index (JI). Mean JI and standard deviation were calculated from the resulting three JI values for each data set.

Ergebnisse/Results:

The JI over all data sets was 0.81 (+- 0.1) indicating highly reproducible volume delineations. The mean segmented volume was 182 ml with a mean deviation of 10 %. In 19 patient data sets volumes outside the presumptive tumour volume were segmented as positive, especially in the cerebellar region. This finding is well comparable to the clinical experience and these volumes were also segmented with high reproducibility.

Schlussfolgerungen/Conclusions:

The results show that the ACO approach delivers reproducible volume delineations when applied to FMISO PET patient data sets. Further development of the proposed algorithm will face excluding positive regions outside the presumptive tumour volume and comparison of automatically generated delineations with manual segmentations by experienced physicians.

Literatur/References:

(1) Lee J.A. (2010) Segmentation of positron emission tomography images: Some recommendations for target delineation in radiation oncology.
Radiotherapy & Oncology, Vol. 96, Issue 3, pp 302-307