A deep neural network for fast and accurate scatter estimation in quantitative SPECT/CT under challenging scatter conditions

Abstract

Purpose: A major challenge for accurate quantitative SPECT imaging of some radionuclides is the inadequacy of simple energy window-based scatter estimation methods, widely available on clinic systems. A deep learning approach for SPECT/CT scatter estimation is investigated as an alternative to computationally expensive Monte Carlo (MC) methods for challenging SPECT radionuclides, such as ⁹⁰Y.

Methods: A deep convolutional neural network (DCNN) was trained to separately estimate each scatter projection from the measured ⁹⁰Y bremsstrahlung SPECT emission projection and CT attenuation projection that form the network inputs. The 13-layer deep architecture consisted of separate paths for the emission and attenuation projection that are concatenated before the final convolution steps. The training label consisted of MC-generated "true" scatter projections in phantoms (MC is needed only for training) with the mean square difference relative to the model output serving as the loss function. The test data set included a simulated sphere phantom with a lung insert, measurements of a liver phantom, and patients after ⁹⁰Y radioembolization. OS-EM SPECT reconstruction without scatter correction (NO-SC), with the true scatter (TRUE-SC) (available for simulated data only), with the DCNN estimated scatter (DCNN-SC), and with a previously developed MC scatter model (MC-SC) were compared, including with ⁹⁰Y PET when available.

Results: The contrast recovery (CR) vs. noise and lung insert residual error vs. noise curves for images reconstructed with DCNN-SC and MC-SC estimates were similar. At the same noise level of 10% (across multiple realizations), the average sphere CR was 24%, 52%, 55%, and 67% for NO-SC, MC-SC, DCNN-SC, and TRUE-SC, respectively. For the liver phantom, the average CR for liver inserts were 32%, 73%, and 65% for NO-SC, MC-SC, and DCNN-SC, respectively while the corresponding values for average contrast-to-noise ratio (visibility index) in low-concentration extra-hepatic inserts were 2, 19, and 61, respectively. In patients, there was high concordance between lesion-to-liver uptake ratios for SPECT reconstruction with DCNN-SC (median 4.8, range 0.02-13.8) compared with MC-SC (median 4.0, range 0.13-12.1; CCC = 0.98) and with ⁹⁰Y PET (median 4.9, range 0.02-11.2; CCC = 0.96) while the concordance with NO-SC was poor (median 2.8, range 0.3-7.2; CCC = 0.59). The trained DCNN took ~ 40 s (using a single i5 processor on a desktop computer) to generate the scatter estimates for all 128 views in a patient scan, compared to ~ 80 min for the MC scatter model using 12 processors.

Conclusions: For diverse ⁹⁰Y test data that included patient studies, we demonstrated comparable performance between images reconstructed with deep learning and MC-based scatter estimates using metrics relevant for dosimetry and for safety. This approach that can be generalized to other radionuclides by changing the training data is well suited for real-time clinical use because of the high speed, orders of magnitude faster than MC, while maintaining high accuracy.

Keywords: 90Y; Convolutional neural network (CNN); Deep learning; Monte Carlo methods; Radioembolization; SPECT/CT; Scatter correction.

Fig. 1

Overview of the scatter correction workflow showing the data generation, training and testing of the trained DCNN.

Fig. 2

DCNN architecture. Number near each block refers to the number of filters for each layer. Each projection view is processed independently.

Fig. 3

Axis slices of the different reconstructions of the test phantom simulation (top row) and profiles across the center. Images have been normalized by the sum counts. The NRMSE images relative to the true phantom are shown in the bottom row.

Fig. 4

Test results from the multiple realizations of the phantom of Fig. 3. (a) Noise vs. contrast recovery averaged over 6 spheres (b) Noise vs. residual count error for the lung insert. The data points represent OS-EM iteration number.

Fig. 5

Measured ⁹⁰Y phantom images. (a) Axiel and coronal SPECT/CT slices of the liver phantom for the different reconstructions. (b) Center slice and profiles of the non-uniform sphere for the different SPECT/CT reconstructions compared with ⁹⁰Y PET. SPECT images without scatter correction show high counts in the regions without ⁹⁰Y and demonstrate the improvement in contrast with MC and DCNN scatter estimation. Images have been normalized to the maximum counts in the slice.

Fig. 6

Comparison of SPECT/CT and PET/CT images following Y-90 radioembolization (a) Patient with large 818 mL lesion with a necrotic center and enhacing rim treated with 3.9 GBq to the left-lobe (b) Patient with a 6 mL lesion treated with 2.9 GBq to the right lobe.

Fig. 7

Patient lesion-to-non-tumoral liver uptake ratios from the different Y90 SPECT/CT reconstructions plotted vs. the corresponding ratios from 90Y PET/CT. The dashed line indicates the identity line. Full results for the 13 lesions in 6 patients are given in Supplemental Table 1.