Dual SwinUNet architecture for enhanced photoacoustic imaging: a contrastive learning approach in image and sinogram domains.

Significance: Leveraging complementary information from multiple data representations enables richer feature extraction and greater robustness under challenging acquisition conditions. Such an approach enhances the accuracy, reliability, and generalizability of photoacoustic image reconstruction, ultimately leading to improved image quality and stronger performance across diverse photoacoustic imaging (PAI) scenarios. Aim: We propose a transformer-based dual SwinUNet architecture that learns features from both the image and sinogram domains to improve PAI within a contrastive learning framework. Approach: The developed dual SwinUNet architecture had multiple loss function-wherein noise-to-signal ratio was computed between the predicted output from each SwinUNet model and the ground truth, and the mean square error was estimated between the predicted outputs of both networks. The dual SwinUNet model was fed with two reconstructed images as inputs, generated using different reconstruction algorithms, i.e., backprojection and Tikhonov-regularized reconstruction. These input pairs can either be positive, meaning both inputs share the same ground truth, or negative, meaning they have different ground truths. The model was trained using a contrastive loss in the sinogram domain, enabling the model to learn distinctive features from both positive and negative pairs. Results: The performance of the different networks (ResNet, UNet, FDUNet, TNet, and the proposed network) was evaluated by varying the number of transducers, angular coverage, and noise levels. The data acquired with 100 transducers having a coverage angle of 135 deg have shown that the structural similarity index measure (SSIM) was improved by 7.5% and the universal image quality index improved by 18% compared with FDUNet. For Conclusions: The dual SwinUNet architecture demonstrates significant improvement in image quality for PAI by learning features from both the image and sinogram domains. The proposed framework can be extended using different DL architectures alongside different analytical/model-based reconstruction inputs.