Application

A deep-learning based method for generating diagnostic-quality PET images from lower quality images to reduce scanning time and radiation exposure

Key Benefits

• Generates a high quality synthetic high-count PET image from low-count PET image or CT scan image
• Reduces radiation exposure to patient
• Reduces scanning time

Market Summary

Positron emission tomography (PET) scans are typically used in evaluating diseases such as cancer, heart disease, and brain disorders. To obtain high quality PET scans, patients are required to adopt multiple bed positions for a sustained length of time, increasing their radiation exposure. There is a continued desire to shorten scanning time to reduce radiation exposure for patients with increased lifetime cancer risk and increase motion control, patient comfort, and scanner throughput. However, shorter scanning time without processing adjustments can result in increased image noise, reduced contrast-to-noise ratio, and degraded image quality. Therefore, further improvements in hardware design and processing are needed to drive down scanning time without adversely affecting image quality and quantification.

Technology Summary

The current invention is a deep-learning based method called cycleGAN, that predicts high-quality full count whole-body PET images from low count (low quality) PET data. CycleGAN architecture improves upon generative adversarial network (GAN) algorithms by including both a forward and reverse training model on the low count and full count PET training data. This approach was implemented and evaluated on a retrospective sample of 22 whole-body PET oncology patient datasets. Low count PET data was generated by histogramming the emission data to one-eighth the scan duration for all bed positions. Statistical analyses showed that cycleGAN produced synthetic PET images with higher quality than those produced with the GAN method and that these synthetic images closely resembled the full-count PET images. If implemented, this PET scan processing algorithm could reduce scanning time and radiation exposure without sacrificing image quality.

Development Stage

This method has undergone evaluation on its ability to generate accurate synthetic PET images and been compared to similar methods.