Researchers have developed a new technique, DeepPET, that uses deep learning to turn PET imaging data into high-quality images at a much faster rate than traditional methods. The team’s findings were detailed in a new study published in Medical Image Analysis.
“The purpose of this research was to implement a deep learning network to overcome two of the major bottlenecks in improved image reconstruction for clinical PET,” wrote lead author Ida Häggström, PhD, department of medical physics at Memorial Sloan Kettering (MSK) Cancer Center in New York City, and colleagues. “These are the lack of an automated means for the optimization of advanced image reconstruction algorithms, and the computational expense associated with these state-of-the art methods.”
DeepPET speeds up the image reconstruction process by “learning” what PET images typically look like and the specific PET scanner’s characteristics. Comparing its effectiveness to other image reconstruction methods, the authors observed that DeepPET was 108 times faster than standard ordered subset expectation maximization and three times faster than filtered back-projection.
Häggström et al. believe these improvements in speed, as well as the technique’s ability to produce images of a higher quality, “should lead to higher patient throughput, as well as more reliable and faster diagnoses and treatment decisions, and thus better care for cancer patients.”
The authors also explained that their technique is not necessarily limited to working with PET imaging data.
“Although this work focuses on PET, the methodology presented is also valid for other types of tomographic data, SPECT and CT being the most relevant examples,” the authors wrote. “SPECT data is even noisier than PET data and has poorer intrinsic resolution making a prime candidate for our approach. CT data is much less noisy than PET, and has higher spatial resolution, also making it a suitable candidate for our approach.”
Speaking to MSK for a blog post about her team’s findings, Häggström noted that “no one has done PET imaging in this way before.” She also detailed why MSK was such a good fit for this research.
“MSK has clinical data that we can use to test this system,” she said. “We also have expert radiologists who can look at these images and interpret what they mean for a diagnosis. By combining that expertise with the state-of-the-art computational resources that are available here, we have a great opportunity to have a direct clinical impact.”
Häggström et al. are currently working to prepare DeepPET for clinical testing.