Abstract / Summary
Abstract Automated brain tumour diagnosis using Magnetic Resonance Imaging (MRI) is essential for rapid oncological intervention, yet conventional deep learning approaches are highly parameter-sensitive and offer limited clinical interpretability. This paper introduces a novel Multi-Agentic Deep Learning Framework for brain tumour detection and classification, developed using the Agent framework. The system orchestrates six specialised Qwen LLM-powered intelligent agents: Data Validation, Preprocessing, Training, Hyperparameter Optimisation, Evaluation, and Explainability/Reporting. To maximise diagnostic precision, we integrate state-of-the-art real-time detectors—YOLOv8, YOLOv11, and YOLO26 — with the Asteroid Satellite-Inspired optimisation (ASIO) metaheuristic to dynamically evolve hyperparameters, including learning rate, batch size, and confidence thresholds. Experimental evaluations conducted on a publicly available brain MRI dataset demonstrate that ASIO optimisation yields substantial performance improvements over baseline methods. Specifically, the ASIO-optimised YOLO26 model achieved a peak mean Average Precision (mAP@50) of 0.9321, a box precision of 0.9651, and a recall of 0.8851, outperforming the baseline YOLO26 detector (mAP@50 of 0.8407), as well as the optimised YOLOv8 (mAP@50 of 0.8451) and YOLOv11 (mAP@50 of 0.7751) variants. Furthermore, the integration of Qwen-powered agents enables automated, natural-language diagnostic reasoning and real-time report generation directly from inference metrics, while achieving a rapid inference latency of less than 15 milliseconds per slice. By unifying automated hyperparameter evolution with LLM-based clinical report generation, the proposed multi-agentic framework minimises manual diagnostic latency and establishes a highly transparent, explainable decision-support tool suitable for real-world clinical neuro-radiology workflows.