A cortically-inspired predictive coding framework for polyp segmentation

Polyp segmentation in colonoscopy remains clinically challenging because existing deep learning models operate as feedforward systems: features are extracted, attention weights are computed, and predictions are generated in a single irreversible pass with no opportunity for revision. We propose GRAFNet, a segmentation framework that introduces three biologically inspired computational mechanisms absent from medical image analysis: (1) learnable orientation filters with centre-surround normalisation that adapt to diagnostically relevant edge structures, (2) parallel retinal pathways with lateral inhibition for competitive feature suppression across scales, and (3) predictive coding with inference-time feedback, which iteratively refines predictions by reconciling high-level hypotheses with low-level observations. These mechanisms are integrated through a closed-loop architecture where cortical feedback generates top-down predictions, computes prediction errors, and updates feature representations during inference, enabling the network to dynamically re-examine ambiguous regions. Across five polyp segmentation benchmarks, GRAFNet achieves state-of-the-art performance (0.929 Dice on CVC-ClinicDB, 0.915 on Kvasir-SEG, and 0.946 on CVC-300), with substantial gains on flat lesions (+ 9.8% over the best competitor) and reduced misclassification of challenging anatomical structures such as haustral folds (− 4.8%). Beyond empirical improvements, the framework provides interpretable insights into how recurrent feedback and competitive interactions improve segmentation under ambiguity.