Deep Learning for Low-Dose PET Imaging

A deep learning algorithm for image quality enhancement of low-dose positron emission tomography scans was successfully implemented in a prospective, multi-center, and multi-institutional study for oncology applications.

Like Comment
Read the paper


Positron emission tomography (PET) is a vital tool in oncology for tumor detection, staging, and therapeutic efficacy evaluation. The diagnostic image quality of PET is proportional to the extent of radiopharmaceutical injected into a patient and the duration for which a patient is imaged. A reduction in either dosage or scan duration degrades image quality and may render the scan non-diagnostic. However, there is substantial interest in a reduction of both factors (dosage and duration), for increasing the speed and throughput of imaging examinations, limiting radiation exposure in pediatric populations, reducing inaccuracies resulting from patient motion, and last but not the least, for increasing patient comfort and access to care. Besides classification of images, deep learning (DL) has recently shown promise to improve the image quality of low-count PET scans. However, there has been no evaluation of the generalizability of such DL models across different hospitals, patient groups, and PET scanners with relevant measures of true efficacy in a clinically relevant population. 

Our Solution:

Our manuscript, entitled “Low-count whole-body PET with deep learning in a multicenter and externally validated study”, describes a real-world clinical validation of deep-learning enhancement of fourfold low-count PET [1]. In a multi-vendor and multi-institutional study consisting of three different academic hospitals whose patients were not included in the deep learning training, we implement a low-count-enhancement technique for PET scans in 50 patients with indications of cancer. To maximize the translational potential of this work, we organized our study as follows:

  1. Three board-certified nuclear medicine radiologists from three different academic institutions blindly evaluated the standard of care PET scans as well as the four-fold low-count enhanced (LCE) scans. These radiologists indicated the presence of hypermetabolic lesions in bone, lymph nodes, liver, lung, muscle, spleen, and brain. A qualitative assessment of diagnostic image quality (DIQ) and overall diagnostic confidence (ODC) was also provided.
  2. Using a separate fourth board-certified nuclear medicine radiologist, we evaluated concordance between quantitative standardized uptake values between the two scans for hypermetabolic lesions and reference regions in the aortic blood pool, gluteus muscle, and liver.

Our Major Findings:

  1. We demonstrate the non-inferiority of the four-fold LCE PET scans in diagnostic image quality and overall diagnostic confidence amongst three blinded readers from three different institutions.
  2. We display tumor detection equivalence between the LCE and standard of care PET scans.
  3. We reveal a high accuracy of quantitative standardized uptake values between the LCE and standard of care PET scans.
  4. We demonstrate how inter-scan variations in lesion depiction between the low-count-enhanced and standard dose PET scans were lower than intra- and inter-reader variations.


An example image of subjects scanned on three different PET scanners with body mass index (BMI) of over 30 is shown below. Traditionally, PET scans are sensitive to varying body habitus and suffer from low signal-to-noise ratio in subjects with higher BMI. We demonstrate how our method improves the diagnostic image quality of low-count PET scans even in high BMI patients.

Novelty, Advantages, and Translational Potential:

  1. This is the first work to prospectively implement a deep learning technique to enhance image quality in a multi-site and multi-vendor environment for PET imaging.
  2. The low generalizability of medical imaging deep learning algorithms to datasets and scanners beyond the ones they have been trained with, has been limiting pervasive applications of such methods. We showed that the proposed image quality enhancement algorithm was successfully generalized to three different academic hospitals, each with a different PET/CT scanner and different PET reconstruction algorithms.
  3. Such a method may have value to either reduce dose or exam duration for PET imaging, depending on the specific clinical need.
  4. The deep learning algorithm can directly be applied to already acquired DICOM images, enabling easy deployment of the technique at hospitals and imaging centers. 

Overall, we believe that this study will help translate the tremendous potential of deep learning in medical imaging into a true clinical setting for simultaneously providing improved value to patients, clinicians, hospitals, and payers.


[1]: Chaudhari, A.S., Mittra, E., Davidzon, G.A., Gulaka, P., Gandhi, H., Brown, A., Zhang, T., Srinivas, S., Gong, E., Zaharchuk, G. and Jadvar, H., 2021. Low-count whole-body PET with deep learning in a multicenter and externally validated study. npj Digital Medicine4(1), pp.1-11.

Akshay Chaudhari

Assistant Professor, Stanford University