Deep learning generation of intermediate projections and Monte Carlo based reconstruction improves 177Lu SPECT images reconstructed with sparse acquired projections

Objectives: The objective was to improve image quality of 177Lu-DOTATATE SPECT images reconstructed from sparse acquired projections by constructing a deep neural network for generation of additional projections. The impact of using Monte Carlo based reconstruction for further improvement of the image quality in sparse acquired projections will also be studied.

Methods: A deep neural network with a 3D U-net structure was constructed. The last layer consists of a rectified linear unit and a pixelwise L2 loss function was used for the stochastic gradient descent algorithm. The deep neural network construction and training was performed with Microsoft’s Cognitive Toolkit (CNTK) 2.6 in Python and the trained model was used in C++. The network was trained to generate the intermediate projections for a sparse set of 30 projections to the original 120 projections. 400 patient SPECT images acquired with 120 projections were used in the training. Every forth projection was used for input to the model and the others for the loss function.SPECT images for quality evaluation were reconstructed with the full 120 projections, 30 projections where every forth projection out of the 120 were used, and 120 projections where 90 were generated from the 30 projections using the trained artificial intelligence (AI) model (30-120 AI), i.e. the deep neural network described above. The SPECT images were reconstructed with a standard iterative reconstruction algorithm with attenuation correction (AC-OSEM) and a Monte Carlo based reconstruction algorithm (SARec-OSEM). Image quality was evaluated quantitatively in SPECT acquisitions of a Jaszczak phantom with 177Lu and visually in SPECT acquisitions of 177Lu-DOTATATE in 15 patients by an experienced nuclear medicine physician.

Results: The phantom images showed that the recovery was slightly decreased but the signal to noise ratio (SNR) was greatly improved by adding the deep learning generated projections to the sparse acquired projections. All SARec-OSEM reconstructions were superior to AC-OSEM. No artefacts such as corelated noise, which are normally introduced when bicubic interpolation of projections are applied, could be detected. The nuclear medicine physician ranked the SARec-OSEM 120 or SARec-OSEM 30-120 AI patient images as showing the best overall quality. The reconstructed patient SPECT images with only 30 projections were evaluated to have too high noise level.

Conclusions: Our study showed that using AI to generate intermediate projections from sparsely sampled SPECT increase the SNR with limited loss in recovery. The AI generator showed improved overall image quality and has the potential to reduce the number of acquired SPECT projections. Furthermore, our in house developed Monte Carlo based reconstruction algorithm SARec-OSEM showed improved image quality in both the quantitative phantom measurements and in the patients. By using both deep learning generated projections and SARec reconstruction, it seems possible to reduce the number of projections substantially, thereby enabling an acquisition reduction by a factor of 2-4 for 177Lu-DOTATATE SPECT images.