GRN+: a simplified generative reinforcement network for tissue layer analysis in 3D ultrasound images for chronic low-back pain

Three-dimensional (3D) ultrasound provides real-time, high-resolution images of soft tissues, making it valuable for pain research. However, identifying different tissues for quantitative analysis is a time-consuming process. We developed and validated GRN+, an innovative concatenated multi-model framework designed to automate precise layer segmentation relying on a limited annotated dataset. The proposed GRN+ framework integrates a ResNet-based generator with a U-Net segmentation model. The generator produces new images and paired masks, guided by the segmentation model, through a process termed Segmentation-guided Enhancement (SGE). During training, the generator’s weights are updated based on the segmentation loss gradient. A multi-stage backpropagation strategy was implemented to address the gradient explosion issue and ensure a stable convergence rate within the integrated model. In the first stage, the segmentation loss is propagated back through both the generator and segmentation model. Optimization focused solely on the segmentation model in the second stage, improving mask prediction accuracy by utilizing the generated images. The GRN+ model was evaluated on a dataset of 69 fully annotated 3D ultrasound scans from 29 subjects, with detailed manual labeling of six tissue layers. Our experiment showed when trained on a dataset with only 5% labeled data, GRN+ significantly outperformed all other semi-supervised learning (SSL) methods (p < 0.05) in terms of Dice coefficient, despite not utilizing unlabeled data for unsupervised training. When trained on fully annotated datasets, GRN+ with SGE achieved a 2.16% higher Dice coefficient while maintaining lower computational costs than other models, including deeplabv3plus, V-Net, and Swin-unet. GRN+ enables accurate tissue layer segmentation while minimizing computational costs and the reliance on annotated data, making it an efficient tool for 3D ultrasound image analysis in cLBP patients.