A Systematic Study of Deep Learning Models and xAI Methods for Region-of-Interest Detection in MRI Scans

ResNet50 outperformed newer transformer-based models for knee MRI analysis, suggesting that established CNN architectures remain highly effective for medical imaging tasks, even when compared to cutting-edge Vision Transformers—an important reminder that newer technology isn't always better for specialized applications.

Explainability methods like Grad-CAM provided "clinically meaningful explanations," indicating that AI systems used in medical diagnosis can be made interpretable enough for doctors to understand why the model identified specific regions, which is crucial for clinical trust and adoption.

The study highlights significant limitations of current datasets: the researchers note that transformer models' underperformance may reflect constraints of the MRNet dataset rather than fundamental limitations, suggesting that responsible AI deployment requires adequate, large-scale training data—a critical consideration before implementing these systems in real clinical settings.

Hybrid approaches combining U-Net and MLP classifiers showed promise for interpretability but sacrificed classification accuracy, demonstrating a real trade-off between explainability and performance that developers must carefully navigate when deploying AI in high-stakes medical contexts.

Manual MRI interpretation suffers from "inter-observer variability," meaning AI systems could potentially reduce diagnostic inconsistency—but only if properly validated and integrated as decision-support tools rather than replacements for radiologist expertise.