Room: Exhibit Hall | Forum 8
Purpose: Positron emission tomography (PET) utilizes radioactive tracers to determine cellular metabolic activity and is an invaluable tool in the field of radiation oncology. Unfortunately, PET suffers from low spatial resolution due to various physical and experimental reasons. We hypothesize that PET spatial resolution can be improved by incorporating information from high resolution computed tomography (CT) scans using deep learning.
Methods: We developed a novel method that generated high resolution PET (hrPET) images from CT images using lung and abdomen CT data. The hrPET images were down-sampled to produce low resolution PET (lrPET) images. We trained a novel densely connected UNet (D-Unet) to reconstruct the hrPET image from a combination of lrPET and CT images using a mean squared error (MSE) loss function. We evaluated our model quantitatively on simulated PET data using MSE and qualitatively on an independent head and neck data set.
Results: After training, the MSE of our model was 0.008, and corresponded to a normalized MSE of roughly 1% for 50,000 training images over 5,000 epochs (1 epoch contained 10 images due to memory constraints). The independent simulated testing set consisting of 130 new images had a normalized MSE of 1.6%. The model was also validated in a different site of the body (head and neck), and qualitatively showed improvements over the standard bicubic interpolation method typically used for clinical purposes.
Conclusion: We demonstrate a novel technique capable of enhancing PET spatial resolution using deep learning. While the initial results look promising, we must further validate our model using phantom and animal studies. We hope to implement our deep learning model as a tool that allows clinicians to better discern gross tumor volume (GTV) in the future.