An incremental algorithm for non-convex AI-enhanced medical image processing

Solving non-convex regularized inverse problems is challenging due to their complex optimization landscapes and the presence of multiple local minima. Despite these difficulties, such models are widely adopted in medical imaging, as they can enhance clinically relevant features beyond standard error minimization. This work introduces a hybrid image processing strategy. The proposed framework, denoted as incDG, embeds neural networks within an incremental model-based scheme to provide effective initializations for a non-convex variational solver, which is subsequently refined through regularized updates. This integration of deep learning tools preserves the interpretability of model-based approaches and ensures greater data consistency. Experimental results on TpV-regularized problems, including image deblurring and tomographic reconstruction on synthetic and real CT data, show that incDG achieves high-quality reconstructions, comparable or superior to those obtained with both model-based and deep learning methods, while exhibiting improved robustness and reducing computational costs. Furthermore, a ground-truth-free training strategy is shown to maintain high performance, supporting the practical applicability of the framework in realistic medical scenarios where reliable reference data are unavailable for specific network training.