A Kinetic Neural Network Approach for Absolute Quantification and Change Detection in Positron Emission Tomography

When dealing with absolute quantification of the tracer kinetics in Positron Emission Tomography (PET), one of the most reliable family of methods are the physiologically-consistent compartmental ODE models [5]. In this work we introduce the kinetic neural Network (kNN), a novel technique for absolute quantification inspired by compartmental models and Neural Networks commonly used in Deep Learning. The ODE model’s explicit solution is computed symbolically with the Python package sympy [3], by using the Pade approximation of the exponential matrix [1]. The model is evaluated voxel-by- voxel with PET data and fitted using the Particle Swarm Optimization (PSO) [2, 4] technique. We applied this method on a set of PET images of the same subjects before and after an intervention. At last, we compared the derived parameter images by using Hostelling’s t-squared statistic to evaluate the regions of the brain affected by the treatment.