Deep Learning Based Approach to Quantification of PET Tracer Uptake in Small Tumors

In Positron Emission Tomography (PET), quantification of tumor radiotracer uptake is mainly performed using standardised uptake value and related methods. However, the accuracy of these metrics is limited by the poor spatial resolution and noise properties of PET images. Therefore, there is a great need for new methods that allow for accurate and reproducible quantification of tumor radiotracer uptake, particularly for small regions. In this work, we propose a deep learning approach to improve quantification of PET tracer uptake in small tumors using a 3D convolutional neural network. The network was trained on simulated images that present 3D shapes with typical tumor tracer uptake distributions ('ground truth distributions'), and the corresponding set of simulated PET images. The network was tested on unseen simulated PET images and was shown to robustly estimate the original radiotracer uptake, yielding improved images both in terms of shape and activity distribution. The same network was successful when applied to 3D tumors acquired from physical phantom PET scans.