Classification of Infiltrated Injections During PET/CT Imaging Applying Deep Learning Technique

Objective: Injected dose infiltration can negatively impact quantitative evaluation of Positron Emission Tomography (PET) data by leading to inaccurate calculation of Standardized Uptake Value (SUV) measurements and limiting bioavailability of the tracer in the patient. Recently developed topical gamma scintillation sensors provide a way to monitor Time Activity Curves (TACs) and determine the presence of activity remaining at the injection site after injecting dose. However, TAC analysis and visual inspection by physician of static PET images differ in many cases which has been a recent research concern. In this work, a deep learning (DL) based classification was implemented to study whether this approach can be a viable solution to classify the injection data based on their quality. Method: A supervised machine learning technique was adopted and TACs obtained from the sensors were fed as input to a neural network. The network was trained to classify two classes of data i.e. good quality and poor quality injections. The performance of the network was tested on the basis of 3-fold cross-validation. Result: The network could label the good quality (92.39% data) and poor quality (7.61% data) injection data with around ~98% and ~86% accuracy respectively with an overall accuracy of ~97%. Conclusion: The objective of this work was to examine the feasibility of implementing a DL approach for PET dose injection quality monitoring.