A Novel Attention-guided Curriculum Learning Approach for Multi-modal Medical Image Segmentation and Disease Classification

Recent advances in medical imaging and artificial intelligence require robust deep learning frameworks capable of processing heterogeneous imaging modalities for accurate segmentation and disease classification. Conventional models often suffer from feature heterogeneity, unstable optimization, and poor generalization, creating the need for intelligent attention-guided and adaptive multi-modal learning approaches.

This study proposes a novel Multi-Modal Attention-Guided Curriculum Learning framework integrating modality-aware attention, spatial-channel feature fusion, and curriculum learning for simultaneous segmentation and disease classification. The framework was evaluated on 3,200 paired MRI and CT patient cases containing pixel-level annotations and disease labels.

Experimental analysis demonstrated that the proposed Multi-Modal Attention-Guided Curriculum Learning framework outperformed Multi-Modal U-Net, Attention U-Net, Swin-UNet, and CNN–Transformer Hybrid models. MAGCL achieved 95.9% classification accuracy, 95.7% F1-score, 95.6% Dice score, 91.2% Intersection over Union, and 0.978 AUC-ROC with improved training stability and generalization.

The superior performance of the proposed Multi-Modal Attention-Guided Curriculum Learning is attributed to the integration of modality-aware attention and progressive curriculum learning. Spatial and channel attention mechanisms improved pathological feature localization, while curriculum learning stabilized optimization through gradual inclusion of difficult samples, resulting in enhanced feature discrimination, segmentation accuracy, and disease classification reliability.

The proposed Multi-Modal Attention-Guided Curriculum Learning framework provides an effective and clinically reliable solution for multi-modal medical image segmentation and disease classification. By integrating modality-aware attention with curriculum learning, the framework significantly improves feature fusion, segmentation precision, disease prediction accuracy, and optimization stability, demonstrating strong potential for intelligent clinical decision-support systems.