GRAFNet: Multiscale Retinal Processing via Guided Cortical Attention Feedback for Enhancing Medical Image Polyp Segmentation
arXiv:2602.15072v1 Announce Type: cross 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 pathw
arXiv:2602.15072v1 Announce Type: cross 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.
Executive Summary
This article proposes GRAFNet, a biologically inspired deep learning architecture for enhancing medical image polyp segmentation. GRAFNet integrates three key modules: a Guided Asymmetric Attention Module (GAAM), a MultiScale Retinal Module (MSRM), and a Guided Cortical Attention Feedback Module (GCAFM). These modules are unified in a Polyp Encoder-Decoder Module (PEDM) that enforces spatial-semantic consistency via resolution-adaptive feedback. Extensive experiments demonstrate consistent state-of-the-art performance on five public benchmarks, with significant improvements over leading methods. GRAFNet establishes a paradigm where neural computation principles bridge the gap between AI accuracy and clinically trustworthy reasoning.
Key Points
- ▸ GRAFNet is a biologically inspired deep learning architecture for medical image polyp segmentation
- ▸ GRAFNet integrates three key modules: GAAM, MSRM, and GCAFM
- ▸ GRAFNet achieves state-of-the-art performance on five public benchmarks
Merits
Strength
GRAFNet's biologically inspired design and multi-scale processing capabilities enable robust polyp segmentation, reducing false positives and false negatives.
Strength
GRAFNet's Guided Cortical Attention Feedback Module (GCAFM) enables iterative refinement and improvement of segmentation results.
Strength
GRAFNet's resolution-adaptive feedback mechanism ensures spatial-semantic consistency in segmentation results.
Demerits
Limitation
GRAFNet's complexity and computational requirements may be a challenge for deployment in resource-constrained clinical settings.
Limitation
GRAFNet's performance may be sensitive to the quality and quantity of training data.
Limitation
GRAFNet's interpretability and explainability may be limited due to the complexity of its neural architecture.
Expert Commentary
GRAFNet's innovative architecture and impressive performance demonstrate the potential of biologically inspired deep learning approaches in medical imaging. However, the model's complexity and sensitivity to training data quality raise concerns about its practical deployment and scalability. To address these limitations, future work should focus on developing more efficient and robust architectures that can be adapted to diverse clinical settings. Furthermore, efforts should be made to improve the model's explainability and interpretability, enabling clinicians to trust and effectively utilize AI-driven segmentation results.
Recommendations
- ✓ Future work should prioritize developing more efficient and robust GRAFNet architectures for practical deployment in resource-constrained clinical settings.
- ✓ Investigations should be conducted to improve GRAFNet's explainability and interpretability, enabling clinicians to trust and effectively utilize AI-driven segmentation results.
