Investigation of lesion detectability using deep learning based denoising methods in oncology PET: a cross-center phantom study

Objectives: Previous studies have demonstrated promising performance in reducing noise in low-statistics PET images when using a deep convolutional neural network (CNN). While existing deep learning-based PET denoising studies have primarily focused on lesion quantification, this work investigates the impact of noise reduction with deep learning on lesion detectability for different noise levels based on a cross-center phantom study.

Methods: For the deep learning network, we used a 3D U-Net trained by the datasets from 70 patients (age: 60.2 ± 22.7 years; BMI: 26.6 ± 4.9 kg/m²) acquired on a Biograph mCT scanner at Yale-New Haven Hospital (YNHH). The list-mode data were obtained for 15.9 ± 3.7 mins after 9.95 ± 0.81 mCi ¹⁸F-FDG injection with continuous-bed-motion whole-body scanning to generate the standard-statistics images using OSEM reconstruction with PSF and TOF modelling. Ten samples of low-statistics images were generated by 10% down-sampling of the list-mode data. Using the low-statistics images as input and the standard-statistics images as label, the network was trained to minimize a L2 loss function with the Adam optimizer. To investigate lesion detectability as a function of count statistics, the SNMMI CTN phantom filled with ¹⁸F-FDG was scanned on a mCT scanner at the University of Pennsylvania. In addition, a uniform cylinder phantom was used to emulate the liver. List-mode data were acquired for 60 minutes. Spherical lesions (9.89 mm diameter) were embedded into the phantoms data by combining the lesion data acquired separately in air with the list-mode data of the two phantoms. Images of six different noise levels (scan duration: 0.5, 1, 1.5, 2, 3 and 6 mins) were generated with 10 replicates for each noise level. The images were reconstructed using OSEM algorithm with PSF and TOF modeling. The denoised images were obtained using the U-Net trained by YNHH data. A generalized scan statistics methodology was used to determine the area under the LROC curve (ALROC) as a function of scan time for images without and with denoising to determine the improved lesion detectability with deep learning based noise reduction.

Results: The noise of the CTN phantom images was effectively reduced by the U-Net for all noise levels. After noise reduction, the mean contrast and max contrast was reduced for all six noise levels in both liver and lung. However, benefiting from the reduced noise in the background, there was a substantial increase in the ALROC values for liver lesions. For lung lesions, the ALROC values derived from images without and with U-Net noise reduction were similar, likely due to the high contrast of the embedded lesions (6:1).

Conclusions: Deep learning based noise reduction methods can improve the detectability of small lesions in low-statistics PET images.

• Download figure
• Open in new tab
• Download powerpoint