An Agentic AI Model for MRI-Based Staging and Grading of Endometrial Cancer using MedSAM and ViT

Endometrial Cancer (EC) is one of the most prevalent gynecological cancers globally. Accurate preoperative staging and the subsequent tumor grading are crucial for prognosis, surgical decisions, and treatment plans. Generally, for MRI evaluation, the radiologists depend totally on their own interpretation, which will be a mixture of subjectivity, a lengthy process, and inconsistency among different observers. The problems outlined above are addressed by creating a highly capable Artificial Intelligence (AI) framework that relies on multiple MRI modalities and clinical attributes for the precise identification of Deep Myometrial Invasion (DMI) and tumor grading. For accurate anatomical segmentation, the framework makes use of the Medical Segment Anything Model (MedSAM). After that, a Uterine Cavity Line Generation Algorithm (UCLGA) is employed to determine the depth of myometrial invasion. The Vision Transformer (ViT) model, fine-tuned via Low-Rank Adoption (LoRA), is used for feature representation and learning. Simultaneously, the agentic reasoning module successively enhances prediction through self-reflection and clinical knowledge. Furthermore, multiple smart agents are applied for segmentation visualization, report generation, compliance monitoring, and scheduling to provide a modular and interpretable system. The model is tested on the EC-MRI dataset, and its overall accuracy reached 96.82%, sensitivity 95.85%, precision 96.77%, F1 score 95.31%, and specificity 96.64%. The model surpassed the other current models in accuracy. Overall, the findings imply that this model is a clinically significant, elucidative, and efficient AI system that could support oncologists and radiologists in the preoperative EC evaluation.