Meta-Transformers: A Hybrid Approach for Medical Image Segmentation With U-Net and Meta-Learning

The increasing demand for accurate and efficient medical image segmentation has driven the development of advanced Artificial Intelligence (AI) models to assist in diagnosing and planning treatments. In this paper, we propose a novel approach that combines U-Net architecture with attention mechanisms and Model-Agnostic Meta-Learning (MAML) to enhance segmentation performance, particularly in scenarios with limited annotated data. By applying this method to the AMOS dataset for multi-organ segmentation in CT and MRI scans, our model significantly improves generalization across different patients and imaging modalities. The use of attention mechanisms refines feature extraction, while class weighting effectively addresses class imbalance. The proposed model achieves a state-of-the-art Dice score of 98.09%, pixel accuracy of 98.39%, precision of 98.58% and recall of 98.27% outperforming existing methods and demonstrating the potential for rapid adaptation in clinical applications. This combination of meta-learning and deep learning offers a robust solution for addressing the challenges of medical image segmentation.