Automatic analysis of molecular PET/CT images using cognition network technology: A novel concept

Objectives Development of an automatic user-independent routine for segregation and quantification (e.g. molecular tumour volume (MTV)) of receptor positive neoplastic lesions in molecular PET/CT DICOM sets.

Methods The routine has been developed in the Cognition Network Language (CNL), generating a semantic object network of context objects (spleen, kidneys, liver, tumoural foci, etc.) using a suitable development environment. Algorithm was trained on ten Ga-68 DOTA-TOC PET/CT DICOM sets. Consecutively, processing time (on an Intel 2GHz client (2GB RAM)), and sensitivity of the routine was compared with the analyses of a trained nuclear medicine physician in another 35 DICOM sets.

Results Except for some extra-growth of the spleen into left kidney bounds in 2 instances, slight under growth right kidney in 3 sets, and misclassification of a paraortic lymph node as intrahepatic lesions in 2 series, the developed CNL script was able to robustly define the body contour, organ borders (i.e. lungs, kidneys, spleen, liver and bladder), and all lesions reported by the specialist. Additionally, in 19 studies with multiple (18-112) lesions, tumor foci not described by the specialist were detected. Each data set analysis took 3:17-5:21 minutes (a gain of 2-8 times comparing to human operator) depending on the lesion burden. The border of each lesion to background, MTV (individual and total tumour burden) and SUV could be determined.

Conclusions The results of this prototype routine built on Cognition Network Technology for PET/CT images provide a proof of applicability of the concept enabling the automatic analysis of lesions. It seems especially promising for shortening the analysis time, improving reproducibility, as well as increasing sensitivity in lesion detection