[Paper Review] GRAFNet: Multiscale Retinal Processing via Guided Cortical Attention Feedback for Enhancing Medical Image Polyp Segmentation

Abdul Joseph Fofanah, Lian Wen+4 more|arXiv (Cornell University)|Feb 15, 2026

Retinal Imaging and Analysis0 citations

TL;DR

GRAFNet introduces a biologically inspired architecture with Guided Asymmetric Attention, Multiscale Retinal Modules, and Guided Cortical Attention Feedback to enhance polyp segmentation, achieving state-of-the-art results and better generalisation across five benchmarks.

ABSTRACT

Accurate polyp segmentation in colonoscopy is essential for cancer prevention but remains challenging due to: (1) high morphological variability (from flat to protruding lesions), (2) strong visual similarity to normal structures such as folds and vessels, and (3) the need for robust multi-scale detection. Existing deep learning approaches suffer from unidirectional processing, weak multi-scale fusion, and the absence of anatomical constraints, often leading to false positives (over-segmentation of normal structures) and false negatives (missed subtle flat lesions). We propose GRAFNet, a biologically inspired architecture that emulates the hierarchical organisation of the human visual system. GRAFNet integrates three key modules: (1) a Guided Asymmetric Attention Module (GAAM) that mimics orientation-tuned cortical neurones to emphasise polyp boundaries, (2) a MultiScale Retinal Module (MSRM) that replicates retinal ganglion cell pathways for parallel multi-feature analysis, and (3) a Guided Cortical Attention Feedback Module (GCAFM) that applies predictive coding for iterative refinement. These are unified in a Polyp Encoder-Decoder Module (PEDM) that enforces spatial-semantic consistency via resolution-adaptive feedback. Extensive experiments on five public benchmarks (Kvasir-SEG, CVC-300, CVC-ColonDB, CVC-Clinic, and PolypGen) demonstrate consistent state-of-the-art performance, with 3-8% Dice improvements and 10-20% higher generalisation over leading methods, while offering interpretable decision pathways. This work establishes a paradigm in which neural computation principles bridge the gap between AI accuracy and clinically trustworthy reasoning. Code is available at https://github.com/afofanah/GRAFNet.

Motivation & Objective

• Motivate accurate polyp segmentation in colonoscopy under diverse morphologies and imaging conditions.
• Develop a biologically plausible architecture that integrates retinal pathways with cortical feedback.
• Enforce spatial–semantic consistency through an encoder–decoder with resolution-adaptive feedback.
• Provide interpretable decision pathways via attention-guided, feedback-driven processing.

Proposed method

• Introduce GAAM to emulate orientation-tuned V1 neurons for boundary-enhanced attention.
• Implement MSRM to replicate retinal parallel pathways (parvocellular, magnocellular, koniocellular, ON–OFF) for multi-feature analysis.
• Add GCAFM to apply predictive coding and refine features with high-level anatomical priors.
• Embed a Polyp Encoder–Decoder Module (PEDM) for hierarchical, resolution-adaptive feedback coordination.
• Train with a bio-inspired loss LBIO combining segmentation loss with feedback consistency and attention guidance terms.

Experimental results

Research questions

• RQ1 RQ1: Does cortical feedback improve segmentation performance versus standard attention and state-of-the-art methods?
• RQ2 RQ2: Do multiscale retinal pathways reduce false positives on normal anatomy?
• RQ3 RQ3: Does asymmetric (orientation-tuned) attention aid detection of subtle flat lesions?
• RQ4 RQ4: Does guided feedback prevent attention drift across scales?
• RQ5 RQ5: What is each biological module’s contribution to performance (ablation)?
• RQ6 RQ6: Does the neurobiological design improve cross-dataset generalisation?

Key findings

• GRAFNet achieves state-of-the-art segmentation on five datasets, with Dice improvements of 3–8% and 10–20% higher generalisation over leading methods.
• On CVC-ClinicDB and Kvasir-SEG, Dice scores reach 0.9290 and 0.9146 respectively, with BF1 around 0.9090 and 0.9163.
• On CVC-ColonDB, CVC-300, the method attains the best or near-best scores across multiple metrics, including Dice and IoU (e.g., Dice up to 0.9461 on some comparisons).
• Ablation shows MSRM provides substantial gains first, followed by GAAM and GCAFM, culminating in the full GRAFNet achieving Dice of 0.9425 on ClinicDB/Kvasir-SEG and 0.9461/0.8896 on CVC-ColonDB/CVC-300 respectively.
• GRAFNet demonstrates reduced false positives on normal anatomy (FPR reductions and higher NPV) and strong attention stability across scales (high AC/SC scores).
• Subtle flat lesions and small polyps benefit from asymmetric attention, with consistent Dice gains for flat (<3 mm) and subtle (3–5 mm) categories.