|
|
 |
| |
|
|
|
Digital Twin-Driven Autonomous Imaging Agent Training for Scalable Diagnostic Access in Underserved Clinical Settings |
|
|
|
PP: 959-981 |
|
|
doi:10.18576/amis/200409
|
|
|
|
Author(s) |
|
|
|
Ahmed I. Taloba,
Omar Mohammad Amer Huwari,
Fuad S. Al-Duais,
Mohammad Kanan,
|
|
|
|
Abstract |
|
|
| Ultrasound imaging plays a critical role in rapid and non-invasive clinical diagnosis; however, existing intelligent diagnostic systems remain limited by domain inconsistency, insufficient annotated datasets, poor cross-domain generalization, and the absence of adaptive autonomous interaction mechanisms, particularly in underserved clinical environments. Conventional convolutional, transformer-based, and segmentation-guided frameworks achieve promising segmentation performance, yet they fail to provide scalable closed-loop refinement, digital twin interaction learning, and continuous autonomous adaptation under heterogeneous imaging conditions. To address these limitations, this study proposes a Digital Twin-Driven Autonomous Imaging Agent framework that integrates a Domain-Adaptive Vision Transformer Autoencoder (DA-ViT-AE) with a Hierarchical ReAct (H-ReAct) autonomous reasoning agent for adaptive ultrasound diagnostic analysis. The primary novelty of this study lies in the integration of digital twin-based sim-to-real interaction learning, autonomous perception-action co-evolution, reasoning-driven iterative refinement, and feedback-guided adaptive segmentation within a unified closed-loop diagnostic ecosystem. The proposed framework enables continuous perception-action interaction, iterative segmentation refinement, feedback-driven adaptation, and cross-domain representation learning within a digital twin ecosystem. The framework was implemented using Python with PyTorch, TensorFlow, OpenCV, MONAI, and reinforcement learning libraries on GPU-enabled environments. Experiments were conducted using the publicly available abdominal ultrasound simulation and segmentation dataset containing 617 real ultrasound scans and 926 synthetic ultrasound scans with multi- organ annotations. The proposed framework achieved a Dice score of 96.52%, precision of 97.72%, recall of 97.20%, and IoU score of 93.41%, outperforming existing state-of-the-art models including U-Net, MSEUnet, and DSEU-net. The proposed framework improved overall segmentation performance by approximately 1.2%–6.8% compared to existing baseline methods through domain- adaptive perception learning and autonomous iterative refinement. Furthermore, the framework demonstrated stable scalability, rapid low-resource adaptation, and reliable cross-domain generalization, establishing an effective and intelligent ecosystem for scalable autonomous diagnostic assistance in resource-constrained healthcare environments. |
|
|
|
|
|
|
|
