|
|
 |
| |
|
|
|
Self-Directing Diagnostic Agents for Adaptive Multi-Modal Medical Imaging with Clinician-Guided Reasoning |
|
|
|
PP: 983-1007 |
|
|
doi:10.18576/amis/200410
|
|
|
|
Author(s) |
|
|
|
Ahmed I. Taloba,
Omar Mohammad Amer Huwari,
Fuad S. Al-Duais,
Mohammad Kanan,
|
|
|
|
Abstract |
|
|
| Medical imaging has become an essential component of modern healthcare for disease diagnosis, lesion localization, and clinical decision support across modalities such as chest X-ray, computed tomography (CT), and histopathology imaging. Although recent deep learning approaches have achieved promising performance, most existing systems rely on fixed and non- adaptive pipelines that cannot dynamically respond to clinical uncertainty, patient variability, or real-time expert feedback. Furthermore, current agentic AI frameworks lack unified dynamic workflow orchestration, uncertainty-aware reasoning, and clinician-guided adaptive refinement mechanisms. To address these limitations, this study proposes a novel Self-Directing Diagnostic Agent with Dynamic Directed Acyclic Graph reasoning (SDDA-DAG) framework for adaptive multi-modal medical image analysis. The proposed method employs a LLaMA-370B Instruct-based self-directing agent to autonomously generate diagnostic workflows and dynamically construct DAG-based computational pipelines integrating DenseNet-121 CNN, Attention U-Net, and Vision Transformer modules. The framework incorporates Monte Carlo Dropout-based uncertainty estimation, retrieval-augmented clinical validation, clinician- in-the-loop refinement, and cost-aware optimization for reliable and clinically interpretable diagnosis. The proposed framework was implemented using Python, PyTorch, MONAI, Hugging Face Transformers, and CUDA acceleration. Experiments were conducted on NIH ChestX-ray14, PhysioNet CT-ICH, and LNCO2 histopathology datasets. The proposed SDDA-DAG framework achieved average diagnostic accuracy of 93.1%, F1-score of 0.93, and AUC of 0.95, outperforming existing agentic AI models by approximately 17.6% accuracy improvement over GPT-4o-agent and more than 40% improvement over several vision-language baselines. The results demonstrate that adaptive DAG orchestration, uncertainty-aware refinement, and clinician-guided reasoning significantly improve diagnostic reliability, calibration stability, and clinical trustworthiness for next-generation intelligent healthcare systems. |
|
|
|
|
|
|
|
