RadioOneAI: a unified deep learning pipeline for MRI brain tumor analysis

Abstract

MRI brain tumor analysis is a crucial yet time-consuming clinical task often affected by inter-observer variability and the limited availability of annotated datasets. This study proposes RadioOneAI, a unified deep learning framework that combines synthetic data generation, tumor classification, detection, and segmentation into a single cross-verified pipeline. The framework consists of four main modules: (1) synthetic MRI generation using Stable Diffusion 1.5, benchmarked against C-GAN and DCGAN using six quantitative evaluation metrics; (2) tumor classification using InceptionResNetV2 enhanced with Coordinate Attention (CA) and dual explainable AI methods (Grad-CAM and SIDU); (3) tumor detection using YOLOv11, improved through CA, ESRGAN super-resolution preprocessing, and synthetic-to-real progressive learning; and (4) pixel-wise tumor segmentation using YOLOv11 instance segmentation with CA and a lightweight mask refinement branch. All modules employ a synthetic-to-real (S2R) training curriculum, where models are first pretrained on Stable-Diffusion-generated MRI images and then fine-tuned on annotated real datasets. Experimental results show that Stable Diffusion 1.5 provides superior image generation performance, achieving an FID score of 101.53, compared with 133.47 for DCGAN and 223.78 for C-GAN. The complete RadioOneAI pipeline achieves 97.8% mAP@50 for tumor detection, 98.63% classification accuracy, and 0.959 mAP@50 for segmentation, while also producing over 40 standardized clinical measurements per case. In addition, a cross-verification layer reconciles outputs from all modules and incorporates uncertainty-aware escalation, supporting safer and more reliable clinical deployment.