Explainable deep learning for stroke lesion analysis: CNN, transformer, hybrid models, and future directions: a systematic review

Abstract Background Stroke is the second leading cause of death and the third leading cause of disability worldwide, affecting approximately 15 million people annually and imposing an economic burden exceeding US$890 billion. Interpretation of brain CT and MRI images is time-consuming, subject to inter-observer variability, and limited by a shortage of neuroradiology experts, particularly in low-resource settings. Deep learning methods have shown strong potential for automated stroke lesion analysis; however, the rapid emergence of convolutional neural networks (CNNs), transformers, and hybrid architectures requires systematic evaluation. Objectives This systematic review summarizes recent advances in CNN-based, transformer-based, and hybrid CNN–transformer models for stroke lesion segmentation and classification using CT and MRI. It also evaluates diagnostic performance, examines explainable artificial intelligence (XAI) methods, identifies barriers to clinical translation, and outlines future research directions. Methods Studies published between 1 January 2015 and 15 January 2026 were identified through searches of IEEE Xplore, PubMed, Scopus, and SpringerLink. Study quality was assessed using the Stroke AI Quality Assessment (SAI-QA) tool, with 48 studies (64.0%) rated as high quality (Cohen’s κ = 0.80). Due to heterogeneity in datasets, methods, and evaluation metrics, meta-analysis was not performed. Results CNN-based models were most commonly used (56%), with U-Net variants achieving Dice scores of 0.59–0.94 for segmentation. Transformer-based models (24%) achieved classification accuracies up to 99.6% and Dice scores above 0.90 by capturing long-range dependencies. Hybrid CNN–transformer models (16%) generally outperformed individual architectures by combining local and global feature learning. Explainable AI methods such as Grad-CAM, SHAP, and LIME were reported in only 11% of studies, and there was no standardized evaluation framework. Key challenges included class imbalance (24%), limited annotated data (20%), poor generalization across institutions (16%), and computational complexity (11%). Emerging approaches such as federated learning (4%), self-supervised learning (5%), and few-shot learning (3%) showed promise for addressing data limitations and privacy concerns. Clinical implications Hybrid CNN–Transformer models show strong potential for clinical stroke diagnosis by combining local feature extraction with global context understanding. However, limited external validation, poor Explainability, and high computational cost remain key barriers to real-world clinical use. Conclusion Hybrid CNN–transformer architectures currently provide the best balance between accuracy and robustness for stroke lesion analysis. However, Explainability remains underdeveloped, and standardized evaluation frameworks are lacking. Future research should focus on multi-center validation, standardized benchmarking, uncertainty-aware XAI, federated learning, and equitable deployment in low-resource settings. This review guides advancing clinically translatable AI systems in stroke imaging.