From Black Box to Clinical Trust: A Conceptual Review of Explainable and Lightweight Deep Learning for Spine Disease Detection and Segmentation.

Abstract:

Deep learning models have made significant contributions to the accurate detection of spinal anomalies. However, persistent challenges related to model transparency, explainability and computational demands continue to hinder the real-time deployment of AI-driven solutions in clinical settings. This review examines contributions from researchers on improving the explainability and transparency of deep learning models for spine image analysis, with the aim of identifying promising approaches and informing future research directions. Specifically, it explores the potential synergy between explainable AI techniques and lightweight models, with the expectation that such integration will yield models that are simultaneously accurate, interpretable, and clinically deployable. Concepts covered include deep learning architectures applied to spine imaging tasks such as classification and segmentation; lightweight model design strategies; and categories of explainability techniques including Grad-CAM, LIME, and attention mechanisms. Constraints to the full clinical adoption of AI solutions in spine imaging are also discussed. Key findings highlight several gaps in the research field. Dataset limitation issue, absence of standardised metrics for evaluating model interpretability, and challenges surrounding clinical acceptability. These gaps point to research direction that calls for greater collaboration between healthcare professionals and AI researchers, to develop spine imaging solutions that are both explainable and practically usable in clinical environments.