Improved low count quantitative SPECT reconstruction with a trained deep learning based regularizer

Aim: Improving low-count SPECT can lead to shorter scans and enable pre-therapy theranostic imaging for treatment planning in therapies with radionuclides such as Lu-177 (208 keV gamma: 10%) and Y-90 (low bremsstrahlung yield) that have low photon yields. In model-based image reconstruction, using more iterations of an unregularized method increases noise, so incorporating regularization into the image reconstruction is desirable to control noise in the low-count setting. Because conventional mathematically designed regularizers have a noise - resolution tradeoff, we designed a regularizer based on learned convolutional operators with the goal of improving noise and quantification at the same time.

Methods: Our proposed framework, BCD-Net, combines deep-learning with physics-based iterative reconstruction and consists of 2 core modules: 1) The image denoising module removes artifacts from an input image using convolutional filters and soft-thresholding. 2) The image reconstruction module performs regularized reconstruction penalizing the difference between unknown image and the clean image that is denoised by denoising module. To train BCD-Net, we used three Lu-177 patient studies with multiple acquisitions on a Symbia SPECT/CT. Patient images are acquired at 1-5 daysafter 7.4 GBq Lu-177 DOTATATE therapy for neuroendocrine tumors. To generate low-count realizations in training dataset, we resampled the post-therapy (high-count) measurement data with Poisson resampling rate of 0.5% to generate pre-therapy count-level (low-count) measurement (assuming that 37 MBq is used for pre-therapy imaging). To test BCD-Net, we used the following Lu-177 phantom and patient studies with acquisitions on a Symbia SPECT/CT: 1) measurement with hot spheres (‘lesions’) in the warm liver of a torso-phantom with a clinically realistic activity distribution, 2) measurement with six ‘hot’ spheres (2,4,8,16,30 and 113 mL) in a ‘warm’ background, and 3) one patient study not used for training. We performed additional short scans on patient to mimic a range of pre-therapy count-levels. We compared between standard unregularized EM algorithm, total variation (TV) conventional regularization, denoising U-Net, and our proposed BCD-Net. We evaluated each reconstruction method based on contrast-to-noise-ratio (CNR), activity-recovery (AR) with quantification using a point source-based calibration, and root mean squared error (RMSE in Bq/ml). We selected the regularization parameter (for TV and BCD-Net) and iteration number (for EM) to obtain the highest CNR.

Results: In phantom studies, visual comparison showed that the proposed BCD-Net reconstruction best matched the true image whereas unregularized EM generates noisy images and TV generates visually unnatural images. Image generated by denoising U-Net showed that the model is over-fitting. Quantitatively, in liver phantom study, BCD-Net significantly improved contrast between hot ‘lesions’ and warm ‘liver’ while reducing noise, thereby improving CNR. BCD-Net improved ‘lesion’ CNR by 21.8/32.9/96.0 (%), AR by 9.2/4.0/35.0 (%), and RMSE by 0.6/9.5/9.3 (%) compared to EM/TV/U-Net. In sphere phantom study, BCD-Net improved CNR by 119.3/176.7/77.5 (%) compared to EM/TV/U-Net. Bias - noise plots showed that BCD-Net decreases noise while improving AR unlike the conventional regularizer and EM. In patient studies, qualitative comparison showed that low-count BCD-Net reconstruction better matched the high-count images compared to other reconstructions.

Conclusions: In clinically realistic evaluations of low-count Lu-177 SPECT, the proposed BCD-Net reconstruction significantly improved contrast-to-noise-ratio and activity quantification compared with standard image reconstruction, conventional regularization method, and denoising U-Net. Further testing for different radionuclides and diverse imaging conditions are ongoing.