Transformer-based multi-organ lesion detection and cross-patient generalization in radiological imaging for clinical diagnostics

Medical imaging plays a critical role in clinical diagnostics by enabling early detection of lesions across multiple organs; however, existing deep learning approaches, particularly convolutional neural networks, remain limited in capturing global contextual relationships and often fail to generalize across diverse patient populations and imaging conditions. Furthermore, most prior studies focus on single-organ analysis, lack inter-organ contextual modeling, and rarely incorporate uncertainty estimation for clinical reliability. To address these limitations, a novel framework named OMNIA-Net (Organ-aware Multi-organ Network with Integrated Attention) is introduced for multi-organ lesion detection and cross-patient generalization. The method employs a Vision Transformer backbone for global feature extraction, enhanced with organ-aware representation learning, cross-organ attention mechanisms, Model-Agnostic Meta-Learning for adaptability, domain-adaptive normalization for reducing distribution shifts, and an uncertainty estimation module for confidence quantification. The framework is implemented using Python with deep learning libraries and trained on AMOS and DeepLesion datasets. Experimental results demonstrate that OMNIA-Net achieves a mean Average Precision of 92.6%, precision of 91.2%, recall of 89.8%, F1-score of 90.5%, and Intersection over Union of 85.7%. Compared to existing models such as Faster R-CNN (84.2% mAP), YOLOv5 (80.6% mAP), and Vision Transformer (87.9% mAP), the proposed approach shows significant improvements in both detection accuracy and localization performance. The results confirm enhanced robustness, reliable uncertainty estimation, and strong generalization across multiple organs and patient distributions, supporting its applicability in real-world clinical diagnostic systems.