Classification of PET/CT Injection Quality Using Deep Learning Techniques and External Radiation Detectors

Introduction: Standardized Uptake Value (SUV) is a widely used parameter in Positron Emission Tomography (PET) for the quantitative evaluation of PET studies. Dose infiltration is one of the several factors which affects accurate measurement of SUV and leads to inaccurate estimations of activity1. Recently developed external gamma detectors provide a way to acquire time activity curves (TACs) that can be analyzed to determine the presence and degree of infiltrations1. The motivation behind this work is to observe whether training a deep neural network with infiltration variations can be used to accurately classify injections. Materials and

Methods: A recent 2016 publication demonstrated the application and feasibility of scintillation sensors (Lucerno Dynamics, LLC) to identify and characterize infiltrations1. Time Activity Curves (TAC) generated from the sensors on the injection arm and the other arm (known as control arm) correspond to the dose activity and baseline activity respectively. The difference between the two TAC curves from injection arm and control arm has been considered as input to the neural network. For the training purpose, anonymized injection data has been collected from the University of Tennessee Medical Center database of Lucerno Dynamics's sensor readings collected as part of an ongoing quality improvement study. As a preprocessing step, the TAC curves for all the injection data were cropped to the same length and normalized to values between 0 and 1. The training of the injection dataset has been performed by a network that follows a similar architecture like LeNET 5 which is a 7 layer Convolutional Neural Network (CNN) proposed by Lecun et. al2. Softmax Cross-entropy with Adam Optimizer3 has been used to train the network. Results: The original dataset had 960 samples in total with a data classification distribution of 869 good injections, 70 minor infiltration, 18 moderate infiltrations, and 3 severe infiltrations. In order to ensure equal number of samples from each class, data has been generated by splitting the dataset into training and test data and then fitting piecewise cubic spline to the available datasets. As a final step, the coefficients of the fitted polynomial have been randomly changed within ± 10% of their range. The training and test split ratio was 80% : 20% and 200 sets of data were used for each class. The training has been performed on the neural network for 3000 epochs. The training accuracy reached 98% and the test accuracy ranged in between 65% to 70%. The plots of accuracy vs. epochs are shown in the attached figure. However, the test accuracy has been observed individually for each class and the model could predict class 1, 2, 3, 4 with 85%, 70%, 52% and 90% accuracy respectively. Conclusions: A combination of deep learning techniques and TACs from external sensors is a novel approach which can play a significant role in classifying infiltrated injections. With the data constraint factor, the highest accuracy achieved was in the range of 70%. However, the system could identify the good injections with 85% accuracy and with 90% accuracy when using only the original data. Future work aims at training the network with sufficient data with improved model to achieve better accuracy.

• Download figure
• Open in new tab
• Download powerpoint