Cross-vender, cross-tracer, and cross-protocol deep transfer learning for attenuation map generation of cardiac SPECT

Abstract

It has been proved feasible to generate attenuation maps (μ-maps) from cardiac SPECT using deep learning. However, this assumed that the training and testing datasets were acquired using the same scanner, tracer, and protocol. We investigated a robust generation of CT-derived μ-maps from cardiac SPECT acquired by different scanners, tracers, and protocols from the training data. We first pre-trained a network using 120 studies injected with ^99mTc-tetrofosmin acquired from a GE 850 SPECT/CT with 360-degree gantry rotation, which was then fine-tuned and tested using 80 studies injected with ^99mTc-sestamibi acquired from a Philips BrightView SPECT/CT with 180-degree gantry rotation. The error between ground-truth and predicted μ-maps by transfer learning was 5.13 ± 7.02%, as compared to 8.24 ± 5.01% by direct transition without fine-tuning and 6.45 ± 5.75% by limited-sample training. The error between ground-truth and reconstructed images with predicted μ-maps by transfer learning was 1.11 ± 1.57%, as compared to 1.72 ± 1.63% by direct transition and 1.68 ± 1.21% by limited-sample training. It is feasible to apply a network pre-trained by a large amount of data from one scanner to data acquired by another scanner using different tracers and protocols, with proper transfer learning.

Keywords: Attenuation map generation; SPECT/CT; myocardial perfusion imaging; transfer learning.

Figure 1.

Sample SPECT emission images and μ-maps from the GE (red box) and Philips (blue box) SPECT/CT scanners (different patients)..

Figure 2.

Strategies of transfer learning using DuRDN. Mode 1: fine-tuning layer $i_1$ and $o_1$; Mode 2: fine-tuning layer $i_1$ and $o_1-o_3$; Mode 3: fine-tuning layer $i_1-i_3$ and $o_1-o_3$; Mode 4: fine-tuning $L_{all}$, all layers.

Figure 3.

Sample predicted μ-maps by Net A, B, and C using DuRDN with both photopeak and scatter-window images input. The generated artifacts by Net A and Net B were denoted by white and red arrows.

Figure 4.

Voxel-wise correlation maps between the predicted and ground-truth μ-maps with both photopeak and scatter-window images as input based on 70 testing cases. The correlation coefficients and R² were listed in each plot for reference.

Figure 5.

Sample SPECT AC images reconstructed with the predicted μ-maps by Net A, B, and C with both photopeak and scatter-window images as input. The AC image from Net A under-estimated the MPI intensities at the inferior and septal (white arrows). The AC image from Net C over-estimated the MPI intensities at the lateral and anterior (yellow arrows) and thus over-corrected the true defects.

Figure 6.

Segment-wise correlation maps of polar maps (top) and Bland-Altman plots of segment PE (bottom) based on 70 testing cases. The correlation coefficients and R² were listed on the correlation maps. The mean value and the 97.5% confidence interval (1.96 standard deviations) were listed in the Bland-Altman plots.

Figure 7.

Sample 2D extent iso-contours and 3D topographical maps of MPI defects compared to clinical normal datasets. Net A and B over-corrected the MPI intensities and thus under-estimated the defect size at the anterior and basal-lateral (yellow arrows in 2D and red arrows in 3D plots).