3D CNN-based method for automatic reorientation of ¹¹C-acetate cardiac PET images using anchor point detection

Background: Interpreting and diagnosing cardiac PET images in the transaxial plane could complicate image assessment and hinder the detection of perfusion defects. Therefore, reorienting cardiac PET images from the transaxial plane to the short-axis plane is essential. Purpose: A convolutional neural network (CNN)-based method for anchor point detection was proposed to enable the automatic reorientation of ¹¹C-acetate cardiac PET images. Methods: A total of 57 subjects who underwent ¹¹C-acetate PET/CT imaging were enrolled in this study. Forty subjects were assigned to the training set, and 17 subjects to the testing set. Three anchor points (the apex of the left ventricle, the center of the left ventricle base and the center of the right ventricle) were manually annotated and used as the gold standard. A 3D CNN incorporating residual modules and fully connected layers was developed to predict the coordinates of three anchor points. A composite loss function was designed to guide the model training. Results: The predicted coordinates demonstrated a significant correlation with the gold standard (ICCs > 0.75, p < 0.05). Across 17 segments, the average normalized root mean square error (NRMSE) was below 0.082, and the average relative difference was less than 8.69%. No significant differences in pharmacokinetic parameters were observed between manual annotation and the proposed method (all p > 0.05). An NRMSE of 0.053 was achieved on the simulated pseudo image. Conclusions: The 3D CNN-based method for anchor point detection demonstrated performance comparable to the manual approach, providing a novel and effective solution for the reorientation of ¹¹C-acetate cardiac PET image.