Automated glioma classification and segmentation using voxel-wise kinetic analysis in dynamic ¹⁸F-FET PET

Objectives Positron emission tomography (PET) with ¹⁸F-fluoroethyltyrosine (FET) has proven to add valuable information for the diagnosis of gliomas, especially when advanced analysis is applied to the dynamic tracer uptake. Yet this manually performed method is time consuming and subject to inter-observer variability. We extended previous work by the development of an automated voxel-wise analysis of dynamic FET PET scans for the detection and classification of tumor tissue into high (WHO-III/IV, HGG) and low grade (WHO-I/II, LGG) gliomas.

Methods The proposed algorithm is based on fitting the time-activity curve (TAC) of each voxel to a pseudo-pharmacokinetic function which is a simplified surrogate for the wash-in and wash-out of the tracer. Subsequently we used multivariate analysis to derive probability maps which address each voxel to an investigated tissue class (brain, vessel, LGG or HGG) for automatic segmentation into volumes of interest. We retrospectively collected dynamic FET recordings (40 min., approx. 180 MBq FET, Siemens ECAT HR+) of patients with LGG (N=16) and HGG (N=26) where tumor grading was confirmed by histology. Classification learning was performed with tissue specific voxel TACs of LGG (N=9), HGG (N=5), brain and vessel (N=14). The remaining data was used to test the accuracy of our method. A digital dynamic phantom consisting of co-centric spheres for LGG, HGG, vessel and brain (radii: 15, 35, 45 and 55 mm) was generated and randomly filled per voxel with TACs from the test patient data.

Results The spatial sensitivity of the segmentation was 89.3% for brain, 92.9% for vessel, 91.3% for HGG and 91.4% for LGG. The specificity was 96.9% for brain, 98.4% for vessel, 99.7% for HGG and 99.9% for LGG. When analyzing the FET PET scans of the test patient group, we found 6 out of 7 LGGs and all 21 HGGs. In no patient included into analysis glioma detection failed entirely.

Conclusions This work shows that an automated assistant system for the detection and classification of gliomas is feasible