Impact of the image spatial sampling on tumour detection in 18F-FDG PET

Objectives Lesion early detection may change the therapy strategy and thus the patient’s outcome. To our knowledge, the impact of images spatial sampling on small lesion detection has not been assessed. The objective of the work is to compare lesion detection performance for images reconstructed either with a 128´128 (most often used in oncology) or a 256´256 grid in the transaxial plane.

Methods 50 18F-FDG PET acquisitions of torso anthropomorphic phantoms were simulated using an analytical simulator with realistic noise [Comtat, 1999]. Each phantom contained 8 tumors of 0.5mm radius at random locations inside three structures. Two datasets of 50 images each were generated and reconstructed using the AWOSEM method with a spatial sampling of either 128´128 grid of 4.1mm (dataset S₁₂₈) or 256´256 grid of 2.05mm (dataset S₂₅₆). Detection performance between S₁₂₈ and S₂₅₆ was compared using an AFROC with 2 sets of 4 observers each. The S₁₂₈ and S₂₅₆ areas Az under the ROC curves were compared using a Student’s t-test.

Results The figure shows the ROC curves for the lesion detection on the S₁₂₈ and S₂₅₆ datasets. Results obtained using S₂₅₆ are significantly better than using S₁₂₈ for the liver (Az₁₂₈=0.54<Az₂₅₆=0.73,p<0.05) and the soft tissue (Az₁₂₈=0.56<Az₂₅₆=0.73,p<0.05) and seem better for the lung (Az₁₂₈=0.76<Az₂₅₆=0.81,p=0.15).

Conclusions The use of a fine spatial sampling (e.g. a transaxial plane 256´256 grid with 2.05mm) significantly improves the lesion detection task performance for small nodules with a low contrast. This work suggests a fine spatial sampling should systematically be used in place of 128x128 grid in the transaxial plane in oncology PET.