DuSFE: Dual-Channel Squeeze-Fusion-Excitation co-attention for cross-modality registration of cardiac SPECT and CT

Abstract

Myocardial perfusion imaging (MPI) using single-photon emission computed tomography (SPECT) is widely applied for the diagnosis of cardiovascular diseases. Attenuation maps (μ-maps) derived from computed tomography (CT) are utilized for attenuation correction (AC) to improve the diagnostic accuracy of cardiac SPECT. However, in clinical practice, SPECT and CT scans are acquired sequentially, potentially inducing misregistration between the two images and further producing AC artifacts. Conventional intensity-based registration methods show poor performance in the cross-modality registration of SPECT and CT-derived μ-maps since the two imaging modalities might present totally different intensity patterns. Deep learning has shown great potential in medical imaging registration. However, existing deep learning strategies for medical image registration encoded the input images by simply concatenating the feature maps of different convolutional layers, which might not fully extract or fuse the input information. In addition, deep-learning-based cross-modality registration of cardiac SPECT and CT-derived μ-maps has not been investigated before. In this paper, we propose a novel Dual-Channel Squeeze-Fusion-Excitation (DuSFE) co-attention module for the cross-modality rigid registration of cardiac SPECT and CT-derived μ-maps. DuSFE is designed based on the co-attention mechanism of two cross-connected input data streams. The channel-wise or spatial features of SPECT and μ-maps are jointly encoded, fused, and recalibrated in the DuSFE module. DuSFE can be flexibly embedded at multiple convolutional layers to enable gradual feature fusion in different spatial dimensions. Our studies using clinical patient MPI studies demonstrated that the DuSFE-embedded neural network generated significantly lower registration errors and more accurate AC SPECT images than existing methods. We also showed that the DuSFE-embedded network did not over-correct or degrade the registration performance of motion-free cases. The source code of this work is available at https://github.com/XiongchaoChen/DuSFE_CrossRegistration.

Keywords: Attenuation Correction; Cardiac SPECT/CT; Cross modality image registration; Deep learning.

Fig. 1.

Schematic of the proposed registration framework. The input $\mu$-map and NAC SPECT image are fed into two cross-connected CNN streams for feature fusion. In each downsampling layer, the two feature maps are first input to two Residual Dense Blocks (RDB) (Zhang et al., 2018) for feature extraction. The two extracted feature maps are then jointly fed into a DuSFE co-attention module for feature fusion and recalibration. As presented in the bottom left box, each DuSFE module consists of a channel-Squeeze-Fusion-Excitation (cSFE) module for channel-wise recalibration, a spatial-Squeeze-Fusion-Excitation (sSFE) module for spatial recalibration, and an input residual connection for local feature fusion. In the last layer, the two extracted feature maps are concatenated and input to a deep registration module followed by a series of fully connected layers to estimate registration parameters.

Fig. 2.

Frameworks of channel-Squeeze-Fusion-Excitation (cSFE) and spatial-Squeeze-Fusion-Excitation (sSFE) modules used in the DuSFE co-attention module shown in Fig. 1. In cSFE, the channel-wise features of the input $F_1$ and $F_2$ are encoded as vectors $V_1$ and $V_2$ and then fused into $V_{fuse}$ via fully connected layers. The fused channel-wise features $V_{fuse}$ e is then applied back to recalibrate the channel-wise weights of $F_1$ and $F_2$. In sSFE, the spatial features of $F_1$ and $F_2$ are encoded as volumes $M_1$ and $M_2$ and then fused into $M_{fuse}$ via convolutional operators. The fused spatial features $M_{fuse}$ are then applied back to recalibrate the spatial weights of $F_1$ and $F_2$.

Fig. 3.

Registered and ground-truth $\mu$-maps (unit: cm⁻¹) in the transverse and coronal view with overlapped SPECT images in the horizontal-long-axis (HLA) and short-axis (SA) view. Error maps by subtracting the ground-truth $\mu$-map from the predicted $\mu$-maps are presented below each image. The registration errors are denoted by white errors.

Fig. 4.

Voxel-wise correlation maps between the attenuation coefficients (unit: cm⁻¹) of registered and ground-truth $\mu$-maps on 100 cases. The red dash line is the reference identity line $(y=x)$. The correlation coefficients (Corr. Coef.) and coefficients of determination $(\left.{\mathrm{R}}^2\right)$ are listed at the top left of each figure.

Fig. 5.

AC SPECT images reconstructed using the registered and ground-truth $\mu$-maps in the HLA and SA view. SPECT voxel values are the myocardial perfusion intensities after volume mean normalization. Error maps by subtracting the ground-truth AC SPECT image from predicted AC SPECT images are presented below each image. The reconstruction errors are denoted by white arrows.

Fig. 6.

Standard 17-segment model polar maps of the reconstructed AC SPECT images using the ground-truth and registered $\mu$-maps. Regions with under-estimated distribution activities are marked with green arrows, and regions with over-estimated activities are marked with white arrows.

Fig. 7.

Segment-wise correlation maps between the polar maps of the registered and ground-truth AC SPECT images on 100 testing cases. The red dash line is the reference identity line (y = x). The correlation coefficients (Corr. Coef.) and coefficients of determination ($\left.{\mathrm{R}}^2\right)$ are listed at the top left of each figure.

Fig. 8.

Average errors of translation (top left), rotation (top right), $\mu$-maps (bottom left), and AC SPECT images (bottom right) on 400 testing cases after registration with different initial motion errors. In total, 7 datasets were tested, and each point on the line refers to a dataset.

Fig. 9.

Registration of rest $\mu$-maps and stress SPECT images. Registered rest $\mu$-maps (unit: cm⁻¹) are shown on the left side. Reconstructed stress AC SPECT images with volume mean normalization are shown on the right side. Error maps of $\mu$-maps or SPECT images are presented below each image. The registration and reconstruction errors are denoted by white arrows.

Fig. 10.

Ablation studies of DuSFE: registered $\mu$-maps (unit: cm⁻¹) by DuSFE (cSFE), DuSFE (sSFE), DuSFE (Max), DuSFE (Conv), and DuSFE (Proposed) with overlapped SPECT images. Error maps by subtracting ground-truth $\mu$-map from predicted $\mu$-maps are presented below each image. The registration errors are denoted with white arrows.

Fig. 11.

Ablation studies of DuSFE: reconstructed AC SPECT images using the registered $\mu$-maps by DuSFE (cSFE), DuSFE (sSFE), DuSFE (Max), DuSFE (Conv), and DuSFE (Proposed). SPECT voxel values are the myocardial perfusion intensities after volume mean normalization. Error maps by subtracting the ground-truth AC SPECT image from predicted AC SPECT images are presented below each image. The reconstruction errors are denoted with white arrows.