Performance of a deep learning enhancement method applied to PET images acquired with a reduced acquisition time.

Abstract

Background: This study aims to evaluate the performance of a deep learning enhancement method in PET images reconstructed with a shorter acquisition time, and different reconstruction algorithms. The impact of the enhancement on clinical decisions was also assessed.

Material and methods: Thirty-seven subjects underwent clinical whole-body [18F]FDG PET/CT exams with an acquisition time of 1.5 min per bed position. PET images were reconstructed with the OSEM algorithm using 66% counts (imitating 1 min/bed acquisition time) and 100% counts (1.5 min/bed). Images reconstructed from 66% counts were subsequently enhanced using the SubtlePET™ (SP) deep-learning-based software, (Subtle Medical, USA) - with two different software versions (SP1 and SP2). Additionally, images obtained with 66% counts were reconstructed with QClear™ (GE, USA) algorithm and enhanced with SP2. Volumes of interest (VOI) of the lesions and reference VOIs in the liver, brain, bladder, and mediastinum were drawn on OSEM images and copied on SP images. Quantitative SUVmax values per VOI of OSEM or QClear™ and AI-enhanced 'shortened' acquisitions were compared.

Results: Two hundred and fifty-two VOIs were identified (37 for each reference region, and 104 for the lesions) for OSEM, SP1, SP2, and QClear™ images AI-enhanced with SP2. SUVmax values on SP1 images were lower than standard OSEM, but on SP2 differences were smaller (average difference for SP1 11.6%, for SP2 -4.5%). For images reconstructed with QClear™, SUVmax values were higher (average +8.9%, median 6.1%, SD 18.9%). For small lesions with SUVmax values range 2.0 to 4.0 decrease of measured SUVmax was much less significant with SP2 (for liver average -6.5%, median -5.6% for lesions average -5.6%, median - 6.0, SD 5.2%) and showed the best correlation with original OSEM. While no artifacts and good general diagnostic confidence were found in AI-enhanced images, SP1, the images were not equal to the original OSEM - some lesions were hard to spot. SP2 produced images with almost the same quality as the original 1.5 min/bed OSEM reconstruction.

Conclusions: The studied deep learning enhancement method can be used to accelerate PET acquisitions without compromising quantitative SUVmax values. AI-based algorithms can enhance the image quality of accelerated PET acquisitions, enabling the dose reduction to the patients and improving the cost-effectiveness of PET/CT imaging.