ML segmentation strategies for object interference compensation in FDG-PET lesion quantification

Quantification of lesion activity by FDG uptake in oncological PET is severely limited by partial volume effects. A maximum likelihood (ML) expectation maximization (EM) algorithm considering regional basis functions (AWOSEM-region) had been previously developed. Regional basis functions are iteratively segmented and quantified, thus identifying the volume and the activity of the lesion.

Improvement of AWOSEM-region when analyzing proximal interfering hot objects is addressed by proper segmentation initialization steps and models of spill-out and partial volume effects. Conditions relevant to lung PET-CT studies are considered: 1) lesion close to hot organ (e.g. chest wall, heart and mediastinum), 2) two close lesions.

CT image was considered for pre-segmenting hot anatomical structures, never for lesion identification, solely defined by iterations on PET data. Further resolution recovery beyond the smooth standard clinical image was necessary to start lesion segmentation. A watershed algorithm was used to separate two close lesions. A subtraction of the spill-out from a nearby hot organ was introduced to enhance a lesion for the initial segmentation and start the further quantification steps. Biograph scanner blurring was modeled from phantom data in order to implement the procedure for 3D clinical lung studies.

In simulations, the procedure was able to separate structures as close as one pixel-size (2.25 mm). Robustness against the input segmentation errors defining the addressed objects was tested showing that convergence was not sensitive to initial volume overestimates up to 130%. Poor robustness was found against underestimates. A clinical study of a small lung lesion close to chest wall displayed a good recovery of both lesion activity and volume.

With proper initialization and models of spill-out from hot organs, AWOSEM-region can be successfully applied to lung oncological studies.