Neurosymbolic Vision Mamba With Hypergraph Reasoning For Anatomically Consistent Diabetic Foot Ulcer Segmentation

Diabetic foot ulcers are chronic wounds whose size and shape must be assessed repeatedly during treatment, and outlining them by hand is slow and varies between observers. This project built and compared four deep learning models that segment ulcer regions in clinical photographs. Each model combines an image encoder, a Mamba inspired gated context block, a hypergraph module over feature nodes and a convolutional decoder. The models differ in the encoder (ResNet 50 and ConvNeXt B with ImageNet weights, and two networks trained from scratch), in how nodes and hyperedges are constructed, in the loss function, and in whether a structural rule about wound bed and wound edge regions is applied. The models were trained under one protocol on 1,370 images pooled from the Foot Ulcer Segmentation Challenge (FUSC) and Medetec collections (913 training, 228 validation and 229 test images) and evaluated with Dice, IoU, precision, recall and HD95, using exponential moving average weights and six fold test time augmentation. On the 168 test images that contain a wound, Model 2 reached a Dice of 0.9367 and Model 1 reached 0.9365, a difference far inside their bootstrap intervals; Model 4 reached 0.9007 and Model 3 reached 0.8759. A review of the code showed that the evaluation is easier than whole photograph segmentation, because images were cropped around the reference mask, and that the 200 FUSC test photographs, whose masks the organisers withhold, were treated as empty masks. No ablation or baseline was run, so the contribution of the individual components is not established. The scores should be read within this protocol. The work is an exploratory architectural study and not a clinical tool.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11718975/v1
Primary Topic
Medical Image Segmentation Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Neurosymbolic Vision Mamba With Hypergraph Reasoning For Anatomically Consistent Diabetic Foot Ulcer Segmentation

Haroon Iqbal, Jan Adnan Farooq, Shayam Ahmad
Research Square
Medical Image Segmentation Techniques
preprint

Neurosymbolic Vision Mamba With Hypergraph Reasoning For Anatomically Consistent Diabetic Foot Ulcer Segmentation

Haroon Iqbal, Jan Adnan Farooq, Shayam Ahmad
preprint en

Abstract

Diabetic foot ulcers are chronic wounds whose size and shape must be assessed repeatedly during treatment, and outlining them by hand is slow and varies between observers. This project built and compared four deep learning models that segment ulcer regions in clinical photographs. Each model combines an image encoder, a Mamba inspired gated context block, a hypergraph module over feature nodes and a convolutional decoder. The models differ in the encoder (ResNet 50 and ConvNeXt B with ImageNet weights, and two networks trained from scratch), in how nodes and hyperedges are constructed, in the loss function, and in whether a structural rule about wound bed and wound edge regions is applied. The models were trained under one protocol on 1,370 images pooled from the Foot Ulcer Segmentation Challenge (FUSC) and Medetec collections (913 training, 228 validation and 229 test images) and evaluated with Dice, IoU, precision, recall and HD95, using exponential moving average weights and six fold test time augmentation. On the 168 test images that contain a wound, Model 2 reached a Dice of 0.9367 and Model 1 reached 0.9365, a difference far inside their bootstrap intervals; Model 4 reached 0.9007 and Model 3 reached 0.8759. A review of the code showed that the evaluation is easier than whole photograph segmentation, because images were cropped around the reference mask, and that the 200 FUSC test photographs, whose masks the organisers withhold, were treated as empty masks. No ablation or baseline was run, so the contribution of the individual components is not established. The scores should be read within this protocol. The work is an exploratory architectural study and not a clinical tool.

Research Square
C.V. Raman Global University (IN)
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon