Visual Mamba-CNN for scribble-based segmentation in weakly supervised learning for photoacoustic tomography

Photoacoustic (PA) imaging is a powerful non-invasive medical imaging technique that combines the high contrast of optical imaging with the deep tissue penetration of ultrasound, offering both structural and functional insights into tissues and organs. Organ-level analysis of photoacoustic tomography (PAT) images enables quantification of specific morphological and functional parameters, making accurate organ segmentation a critical step in PA image-based analysis. However, the limited availability of large-scale annotated datasets remains a major challenge. To address this, we employ cross-modality data augmentation by generating synthetic PA images from MRI scans. To further reduce manual annotation efforts, we propose a weakly supervised learning (WSL) framework that leverages scribble annotations. Since many deep learning models struggle to capture global context from sparse labels, we introduce a novel architecture that combines traditional convolutional neural networks (CNNs) with Visual Mamba, integrating both local and global feature extraction capabilities. This hybrid design improves segmentation performance in weakly supervised settings. We validate our method on a simulated PA abdominal dataset and real in vivo mouse abdominal PAT data, demonstrating notable improvements in segmentation accuracy and robustness.