LINDA: Low Resource Foot Ulcer Segmentation using Local Similarity and Iterative Neural Dynamics

Trishaani Acharjee; Sudip Chakrabarty; Rajdeep Chatterjee; Tathagata Bandyopadhyay

cs.CVNov 3, 2026

LINDA: Low Resource Foot Ulcer Segmentation using Local Similarity and Iterative Neural Dynamics

Authors: Trishaani Acharjee, Sudip Chakrabarty, Rajdeep Chatterjee, Tathagata Bandyopadhyay

Organizations: School of Computer Engineering, Kalinga Institute of Industrial Technology, Bhubaneswar, India · AmygdalaAI-India Lab, Bhubaneswar, India · Department of Informatics, Technical University of Munich, Munich, Germany

Abstract

Foot ulcers are a leading cause of lower limb amputation, and the highest burden often falls on low and middle income regions where specialists and computing hardware are both scarce. Wound monitoring tools are most useful there when they are accurate yet small enough to run on ordinary devices, but accuracy usually comes from heavy networks. We present LINDA, a compact model that grounds segmentation in an explicit structural prior instead of raw intensity. A Local Self-Similarity (LSS) map, computed directly from the image and stable under local lighting changes, is fused into a ResNet-34 encoder, and a small Liquid Time-Constant module refines the boundary over four iterative steps. With every model trained under one protocol on the FUSeg 2021 dataset, LINDA obtains the best Dice (86.68%), IoU (79.67%), and boundary accuracy (HD95 6.37 pixels) in our comparison while using only 25.70M parameters, fewer than the U-Net and transformer baselines. As the LSS map is a fixed function of the input, it exposes the structural cues the model follows in a readable form. Together this makes LINDA an accurate, inspectable, and lightweight option for wound assessment in resource-constrained care.

Figures & tables

Figure 1: The proposed LINDA, illustrating the integration of the ResNet-34 encoder, Additive LSS Fusion, LTC Refinement Loop, and Deep Supervision Decoder.

Figure 2: Sample images and ground-truth masks from the FUSeg 2021 challenge dataset.

Method	Params (M)	FLOPs (G)	Dice (%) ( $\uparrow$ )	IoU (%) ( $\uparrow$ )	HD95 ( $\downarrow$ )
VGG16-UNet [ 23 ]	69.31	42.19	84.88	77.29	14.72
TransUNet [ 6 ]	130.22	102.00	80.16	71.32	15.32
UNet [ 5 ]	31.78	374.34	84.04	76.23	13.31
Mask R-CNN [ 24 ]	90.44	67.13	83.97	76.42	12.89
ResNet101-UNet [ 25 ]	267.62	112.79	84.92	77.73	13.03
U2Seg [ 26 ]	86.47	88.46	75.85	70.13	16.07

Table 1: Quantitative comparison of LINDA against representative segmentation methods under a unified training protocol.

Model Variation	LSS	LTC	BAL	Dice (%)( $\uparrow$ )	IoU (%)( $\uparrow$ )	Params (M)
Baseline (ResNet-34 + U-Net)				$85.22$	$77.56$	24.87
Baseline + LSS	✓			$83.51$	$75.51$	25.12
Baseline + LTC		✓		$83.29$	$75.22$	25.45
Baseline + LSS + BAL	✓		✓	$86.41$	$79.26$	25.12
LSS + Decoder only + BAL	✓		✓	$46.09$	$34.61$	3.01
LINDA (w/o BAL)	✓	✓		$76.18$	$69.74$	25.70

Table 2: Ablation study progressively adding LSS, LTC, and BAL to the baseline.

T	Dice ( $\uparrow$ )	IoU ( $\uparrow$ )	HD95 ( $\downarrow$ )	T	Dice ( $\uparrow$ )	IoU ( $\uparrow$ )	HD95 ( $\downarrow$ )
1	86.28	79.33	6.93	8	86.26	79.33	6.64
2	85.27	78.10	11.36	9	69.21	57.54	30.91
3	84.66	77.12	9.71	10	67.22	55.67	31.02
4	86.68	79.67	6.37	11	72.47	62.35	22.92
5	85.55	78.31	9.34	16	69.79	58.55	27.23
6	86.65	79.57	6.59	20	48.72	41.26	48.35

Table 3: Effect of the number of LTC refinement steps

T

on the FUSeg validation set. Dice/IoU in %, HD95 in pixels;

T{=}4

(shaded) is used.

Figure 3: The three LSS channels (Mean, Max, and Std) for sample wounds. The Std channel highlights the wound rim, while Mean and Max capture local homogeneity and strong local structure.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 822 papers

Explainable Continuous-Time Mask Refinement with Local Self-Similarity Priors for Medical Image Segmentation
Feb 28, 2026Rajdeep Chatterjee, Sudip Chakrabarty, Trishaani AcharjeeMedical Image Segmentation Boundary-Aware Segmentation
School of Computer Engineering, Kalinga Institute of Industrial Technology, Bhubaneswar, India · AmygdalaAI-India Lab, Bhubaneswar, India
Do Diabetic Foot Ulcer Segmentation Models Generalize? A Cross-Dataset Benchmark of CNN and Transformer Architectures
Jun 27, 2026Abderrahmane BenfatahMedical Diagnosis Medical Image Segmentation
King Saud University, Riyadh, Saudi Arabia
Automated multi-class wound assessment using dedicated instance segmentation models for boundary detection and classification
Mar 28, 2026Mehedi Hasan Tusar, Fateme Fayyazbakhsh, Igor Melnychuk +1Instance Segmentation Medical Image Classification
Department of Mechanical and Aerospace Engineering, College of Engineering and Computing, Missouri University of Science and Technology, Rolla, Missouri, United States of America · Missouri Protoplex, Missouri University of Science and Technology, Rolla, Missouri, United States of America · Center for Biomedical Research, Missouri University of Science and Technology, Rolla, Missouri, United States of America +2

Feb 28, 2026J/K move · Enter open · S save

Explainable Continuous-Time Mask Refinement with Local Self-Similarity Priors for Medical Image Segmentation

Rajdeep Chatterjee, Sudip Chakrabarty, Trishaani Acharjee

School of Computer Engineering, Kalinga Institute of Industrial Technology, Bhubaneswar, India · AmygdalaAI-India Lab, Bhubaneswar, India

Medical Image Segmentation Boundary-Aware Segmentation Explainable Medical Image Analysis

Open paper AskPDF arXiv

Accurate semantic segmentation of foot ulcers is essential for automated wound monitoring, yet boundary delineation remains challenging due to tissue heterogeneity and poor contrast with surrounding skin. To overcome the limitations of standard intensity-based networks, we present LSS-LTCNet:an ante-hoc explainable framework synergizing deterministic structural priors with continuous-time neural dynamics. Our architecture departs from traditional black-box models by employing a Local Self-Similarity (LSS) mechanism that extracts dense, illumination-invariant texture descriptors to explicitly disentangle necrotic tissue from background artifacts. To enforce topological precision, we introduce a Liquid Time-Constant (LTC) refinement module that treats boundary evolution as an ODEgoverned dynamic system, iteratively refining masks over continuous time-steps. Comprehensive evaluation on the MICCAI FUSeg dataset demonstrates that LSS-LTCNet achieves state-of-the-art boundary alignment, securing a peak Dice score of 86.96% and an exceptional 95th percentile Hausdorff Distance (HD95) of 8.91 pixels. Requiring merely 25.70M parameters, the model significantly outperforms heavier U-Net and transformer baselines in efficiency. By providing inherent visual audit trails alongside high-fidelity predictions, LSS-LTCNet offers a robust and transparent solution for computer-aided diagnosis in mobile healthcare (mHealth) settings.

Abstract

Medical Image Segmentation

Explore similar work

Explainable Continuous-Time Mask Refinement with Local Self-Similarity Priors for Medical Image Segmentation

Do Diabetic Foot Ulcer Segmentation Models Generalize? A Cross-Dataset Benchmark of CNN and Transformer Architectures

Automated multi-class wound assessment using dedicated instance segmentation models for boundary detection and classification