Integrating U-Net in a LLM Supervisor Agent Pipeline for Pancreatic Ductal Adenocarcinoma Diagnosis

Background: Pancreatic Ductal Adenocarcinoma (PDAC) is one of the most common and lethal forms of pancreatic cancer, with a five-year survival rate around 13%. Early detection plays a pivotal role in improving patient prognosis, but current diagnostic workflows can be resource-intensive, increasing the likelihood of missing subtle yet clinically relevant information.There is a critical need for systems that can not only identify clinical features in imaging data but also streamline their translation into actionable insights.We propose an AI-based diagnostic pipeline that integrates U-Net segmentation with a Supervisor Agentic framework.This system is designed to assist in the analysis of CT scans, extracting essential clinical data and providing digestible information for clinicians to use in decision-making.Objective: This work presents a modular, agentic AI pipeline designed to support PDAC diagnosis by automating each stage of the clinical imaging workflow.The system compares and integrates U-Net-based segmentation with an agentic AI pipeline to interpret CT scans, extract clinically relevant features and generate structured summaries for clinician review.By adopting a plug-and-play multi-agent architecture, this pipeline not only enhances diagnostic accuracy and consistency in PDAC care, but also lays the groundwork for scalable, transferable AI solutions adaptable across other diagnostic domains in clinical medicine. Methods:We developed the Supervisor Agent pipeline composed of separate agents for data preparation, model selection, and documentation.A large language model is integrated using the LangChain framework, enabling orchestration of complex, multi-step tasks by chaining language model prompts with external tools.For PDAC validation, six U-Net model variants, including Attention U-Net and TransU-Net, were evaluated using a public dataset from the Decathlon challenge, obtained through The Cancer Imaging Archive (TCIA). Results:The agentic pipeline demonstrated the ability to autonomously manage and interpret radiological workflows.The LLM was fine-tuned and prompted to coordinate agent tasks such as data preprocessing, CT slice prediction, and clinician-readable summary generation to highlight anomalies.Among the segmentation models, TransU-Net achieved the best performance, with a Dice coefficient of 0.89 for pancreatic tissue and 0.85 for pancreatic cancer segmentation.Conclusion: These initial findings support the viability of our agent-driven pipeline in assisting PDAC detection and clinical reporting.By integrating the segmentation models, decision making support and documentation into modular components, the system offers a scalable framework for expanding AI assistance in diagnosis.Future integration of a retrieval-augmented generation (RAG) agent will enable the system to recommend treatment options based on prior outcomes.