Global and local information-based prostate image registration of prostate-specific membrane antigen PET/CT and enhanced MRI.

[BACKGROUND] Prostate-specific membrane antigen positron emission tomography/computed tomography (PSMA PET/CT) and contrast-enhanced magnetic resonance imaging (MRI) are extensively used for the clinical diagnosis of prostate cancer (PCa). However, both modalities are prone to misdiagnoses or missed lesions. Fusing these complementary data sources, specifically the combined CT-PET image data from PSMA PET/CT scans and contrast-enhanced MRI, may improve diagnostic accuracy, with precise spatial registration being a crucial prerequisite for effective image fusion. While previous studies have used software tools for MRI-PSMA PET/CT fusion, most rely on MRI-CT-based anatomical registration and treat PET as a secondary overlay, thereby underutilizing PSMA's tumor-specific metabolic information.

[PURPOSE] To propose a global and local information-based registration network (GLNet) that integrates PSMA PET/CT's functional-semantic features with MRI's high-resolution soft-tissue details to improve PCa lesion diagnosis.

[METHODS] To improve prostate image registration, GLNet was designed using semantic gating convolutional (SGC) modules and a convolutional long short-term memory network based on a U-shaped channel (U-CLSTM). Specifically, SGC modules enhance perception of the prostate gland using global information, while U-CLSTM improves attention to local tumor regions. The dataset comprised 77 clinical cases, each verified by two experienced physicians through clinical biopsy. After data augmentation, 244 cases were used for training and validation, and 64 cases for testing. GLNet's performance was compared against state-of-the-art methods: symmetric normalization (SyN), VoxelMorph (VM), volume tweening network (VTN), and recursive deformable pyramid (RDP). Statistical analyses were conducted using the Kruskal-Wallis test with Bonferroni correction for pairwise comparisons, and effect sizes were assessed using Cohen's d.

[RESULTS] GLNet achieved Dice similarity coefficient (DSC) of 0.76 0.11, HD95 of 9.32 3.13 mm, average symmetric surface distance (ASD) of 1.69 0.65 mm, and a near-zero negative Jacobian proportion of for prostate gland registration. In contrast to other networks, GLNet demonstrated significant improvements (p 0.001), with DSC increasing by - , 5th-percentile Hausdorff distance (HD95) decreasing by 1.69-30.03 mm, and ASD reducing by 0.24-15.44 mm. Cohen's d values indicated large effect sizes. For local lesion detection, GLNet achieved precision, recall, and an F1-score of 0.86, which outperformed other methods by - in precision, - in recall, and 0.21-0.37 F1-score.

[CONCLUSIONS] This study presents GLNet, a deep learning-based non-rigid registration network that fuses PSMA PET/CT and contrast-enhanced MRI. GLNet outperforms existing methods in registration accuracy and lesion detection, thereby offering a promising approach for integrating structural and functional imaging in clinical PCa diagnosis.