FOCUS-Med: Polyp Segmentation Framework

• FOCUS-Med is a deep learning framework for endoscopic polyp segmentation that fuses graph-based spatial modeling, stand-alone self-attention, and multi-scale feature aggregation.
• It employs a Dual-GCN module, location-fused self-attention, and weighted fast normalized fusion to enhance boundary preservation and improve polyp localization.
• The integration of an LLM-based qualitative evaluation protocol distinguishes FOCUS-Med, achieving state-of-the-art performance on multiple public benchmarks.

Searching arXiv for the specified paper and closely related work on endoscopic polyp segmentation and attention/graph-based methods. arxiv_search(query="(Fan et al., 9 Aug 2025) OR \"LLM Evaluated Stand-alone Attention-Assisted Graph Neural Network with Spatial and Structural Information Interaction for Precise Endoscopic Image Segmentation\" OR FOCUS-Med polyp segmentation", max_results=10) FOCUS-Med is a deep learning framework for precise endoscopic polyp segmentation that integrates graph-based spatial-structural modeling, stand-alone self-attention, and multi-scale feature fusion to address low contrast with surrounding mucosa, specular highlights, and indistinct boundaries in endoscopic images. The method combines a Dual Graph Convolutional Network (Dual-GCN) module, a location-fused stand-alone self-attention mechanism, and a trainable weighted fast normalized fusion strategy, and it additionally introduces a LLM-based qualitative evaluation protocol for segmentation quality. On public benchmarks, it is reported to achieve state-of-the-art performance across five key metrics, with both quantitative and qualitative analyses emphasizing sharper boundaries, improved localization, and clinically oriented assessment (Fan et al., 9 Aug 2025).

1. Clinical task and design objectives

FOCUS-Med targets polyp segmentation in endoscopic imaging, a task described as critical for early colorectal cancer detection. The motivating technical difficulties are low contrast with surrounding mucosa, specular highlights, and indistinct boundaries, all of which complicate precise delineation of polyp extent. Within this problem setting, the framework is designed to capture not only appearance cues from image intensities but also contextual spatial relations and topological structural dependencies that may be obscured in raw images (Fan et al., 9 Aug 2025).

The architecture is organized around three interacting design goals. First, it seeks to preserve boundaries and delineate complex shapes typical of polyps through graph-based reasoning over spatial and structural relationships. Second, it strengthens global context integration with location-aware self-attention. Third, it bridges the semantic gap between encoder and decoder representations through efficient multi-scale aggregation. A further distinguishing feature is that the framework introduces an LLM to provide detailed qualitative evaluations of segmentation quality, rather than limiting assessment to conventional overlap and error metrics (Fan et al., 9 Aug 2025).

A common misunderstanding would be to treat the LLM component as part of the segmentation model itself. The reported pipeline instead uses graph, attention, and fusion modules for mask prediction, while the LLM is used to assess the quality of predicted masks against expert annotations.

2. Architectural organization

The general pipeline begins with an encoder that utilizes ConvNeXt blocks to extract hierarchical image features. At the deepest encoder level, a Dual-GCN Based Feature Enhancement Block (DBFEB) augments these features using both spatial and structural graphs. The decoder then incorporates a location-fused stand-alone self-attention (LFSA) module to refine representations, and a weighted fast normalized fusion (WFNF) strategy combines features from skip connections, the DBFEB, and the decoder before producing the final segmentation mask (Fan et al., 9 Aug 2025).

The architecture can be summarized as follows.

Component	Function	Reported role
ConvNeXt encoder	Hierarchical feature extraction	Produces multi-level image features
DBFEB with Dual-GCN	Bottleneck enhancement	Models spatial and topological dependencies
LFSA decoder module	Contextual refinement	Adds location-aware global context
WFNF	Multi-scale aggregation	Bridges encoder-decoder semantic gaps

This organization reflects a division of labor across representational scales. The encoder supplies hierarchical features, the bottleneck enriches them with graph-structured reasoning, and the decoder restores dense spatial predictions while integrating long-range context and skip-connected detail. This suggests that FOCUS-Med is not merely an attention-augmented encoder-decoder, but a system in which graph construction and attention-based refinement are both treated as primary mechanisms for disambiguating difficult boundaries.

3. Dual-GCN and graph-based feature enhancement

The Dual-GCN module is the central mechanism for modeling spatial and structural information. Its stated purpose is to enhance feature representations by jointly encoding local spatial relations and global structural or topological dependencies. In the spatial branch, each pixel or feature-map position is treated as a node, and edges are weighted by learned similarity or attention between features. Given an input feature tensor $A \in \mathbb{R}^{C \times H \times W}$, the module applies $1 \times 1$ convolutions to generate projections $B$, $C$, and $D$, computes an attention map by $\mathrm{softmax}(B^\top C)$, and forms an enhanced feature as

$$\hat{A} = \mathrm{ReLU}\left( \left( \frac{\exp(B^\top C)}{\sum_{i=1}^{N} \exp(B^\top C)} \right) D + A \right).$$

This branch is described as focusing the network on spatially consistent structures and as being helpful for identifying small, subtle polyps (Fan et al., 9 Aug 2025).

The structural branch addresses longer-range anatomical context and explicit topological priors. It constructs a graph whose edge weights are derived from attention, then applies Dijkstra’s shortest path algorithm to identify salient relations between potentially distant nodes. For nodes $i$ and $j$ with attention scores $\alpha_{ij}^{(k)}$ across $1 \times 1$0 heads, the edge weight is

$1 \times 1$1

Stronger attention is treated as lower path cost. The method considers paths up to a limiting length $1 \times 1$2, filters top-$1 \times 1$3 shortest paths, and performs multi-level attention-based aggregation over both paths and path lengths. The reported node update is

$1 \times 1$4

where $1 \times 1$5 denotes the path-level feature at path length $1 \times 1$6, and $1 \times 1$7 denotes concatenation (Fan et al., 9 Aug 2025).

The stated impact of this design is that the network can reason about polyp boundaries and morphology even in low-contrast, cluttered backgrounds by leveraging both direct and indirect interactions. In effect, the spatial branch privileges local coherence, while the structural branch introduces a path-based notion of global connectivity.

4. Location-fused self-attention and weighted fast normalized fusion

The LFSA module is introduced to capture global dependencies and spatial variability associated with irregular and diverse polyp shapes. Unlike content-only self-attention, LFSA injects explicit spatial location information in the form of relative row and column offsets. Its reported formulation is

$1 \times 1$8

where $1 \times 1$9, $B$0, and $B$1 are query, key, and value representations, $B$2 encodes spatial offsets, $B$3 and $B$4 are trainable parameters, and $B$5 is a regularizer. The module is presented as alleviating boundary uncertainty and enabling location-dependent contextual interactions (Fan et al., 9 Aug 2025).

To aggregate features from multiple sources efficiently, FOCUS-Med uses WFNF. The stated problem is how to combine skip-connected encoder features, decoder features, and DBFEB outputs without excessive computational cost or semantic inconsistency. The reported fusion is a trainable normalized weighted sum:

$B$6

where $B$7 is the fused output, $B$8 are input features, $B$9 are trainable weights, and $C$0 is a small regularization constant (Fan et al., 9 Aug 2025).

Taken together, LFSA and WFNF define the decoder-side strategy of the framework: attention contributes location-aware global context, while fusion adaptively reconciles heterogeneous feature streams across scales. This suggests that the model treats boundary precision not as a purely local decoding issue, but as an outcome of controlled interaction among graph-enhanced, skip-connected, and context-refined representations.

5. Quantitative evaluation and empirical profile

FOCUS-Med was evaluated on four public benchmarks: Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and Endoscene. The reported quantitative metrics are Dice, IoU, MAE, Boundary F-score, and S-measure. The paper states that the method achieves state-of-the-art performance across these five key metrics and reports statistical significance using Wilcoxon signed-rank tests with all $C$1 (Fan et al., 9 Aug 2025).

Representative numerical results are reported for several datasets.

Dataset	Reported FOCUS-Med result	Comparator note
CVC-ClinicDB	Dice $C$2, IoU $C$3	Surpasses DCR-Net Dice $C$4
CVC-ColonDB	Dice $C$5	DCR-Net Dice $C$6
Endoscene	Dice $C$7	DCR-Net Dice $C$8

The qualitative findings are reported as consistent with the quantitative results. Output masks are described as having sharper boundaries, superior polyp localization, and better shape congruence, and feature-map visualizations show higher activation near correct polyp regions. Ablation studies indicate that removing DBFEB, LFSA, or WFNF leads to consistent performance drops, which is presented as evidence that each component contributes materially to the final performance (Fan et al., 9 Aug 2025).

The efficiency profile is also reported. Despite a parameter count of 43M, FOCUS-Med is said to maintain practical inference speed at 94 ms per image, corresponding to 10.6 FPS. This places the model in a regime where architectural complexity is explicitly balanced against deployable throughput rather than treated as a purely offline research cost.

6. LLM-based qualitative evaluation and position in the literature

A distinctive element of FOCUS-Med is its use of an LLM, specifically GPT-4o, for qualitative assessment of segmentation results. For each test sample, the LLM receives the original endoscopic image, the predicted polyp mask, and the expert ground-truth mask, and is prompted to rate five criteria on a 5-point Likert scale: Overall Segmentation Quality (OSQ), Boundary Accuracy (BA), Localization Precision (LP), Shape Consistency (SC), and Clinical Usefulness (CU). The reported findings, based on violin plots and Table 3, are that FOCUS-Med scored higher than the state-of-the-art competitor DCR-Net across all five criteria and exhibited tighter distributions, with interquartile range of 4–5 (Fan et al., 9 Aug 2025).

This evaluation protocol is presented as a human-aligned, multi-dimensional complement to Dice- or IoU-centric assessment. A plausible implication is that the framework attempts to connect algorithmic segmentation quality with clinically legible criteria, particularly boundary fidelity and practical usefulness. At the same time, the LLM is not presented as replacing expert annotation; rather, it operates on predictions already referenced to expert ground truth.

Within the broader segmentation literature represented in the provided corpus, FOCUS-Med belongs to a class of methods that combine local detail modeling with broader contextual reasoning, but it does so through a specific combination of graph reasoning, location-aware stand-alone attention, and LLM-based qualitative scoring. FocusNet emphasizes local and pooling attention for transformer-enhanced polyp segmentation (Zeng et al., 18 Apr 2025), Focal-UNet uses focal modulation to balance local and global features in medical segmentation (Naderi et al., 2022), and FM-BFF-Net combines focal modulation with bidirectional feature fusion across multiple segmentation domains (Safdar et al., 23 Oct 2025). FOCUS-Med differs in centering Dual-GCN-based spatial-structural interaction and in making LLM-mediated qualitative evaluation an explicit part of the reported framework (Fan et al., 9 Aug 2025).