WILLIE: A Unified Framework and Benchmark for Wound Classification, Segmentation, and Localization

Gopi Trinadh Maddikunta; Shannan Hamlin; Hsin-Mei Chen; Kimaya Barnes; Peizhu Qian

cs.CVOct 4, 2026

WILLIE: A Unified Framework and Benchmark for Wound Classification, Segmentation, and Localization

Authors: Gopi Trinadh Maddikunta, Shannan Hamlin, Hsin-Mei Chen, Kimaya Barnes, Peizhu Qian

Organizations: Department of Computer Science University of Houston Houston, Texas, USA · Houston Methodist Academic Institute Houston Methodist Hospital System Houston, Texas, USA

Abstract

Chronic wound management affects over 8.2 million patients in the United States and imposes substantial clinical and economic burden. Clinical wound assessment commonly involves three coupled tasks: identifying wound type, delineating wound boundaries, and localizing the wound region for measurement and monitoring. Despite this clinical coupling, existing machine learning approaches typically address wound classification, segmentation, and localization using separate models. We present WILLIE, a unified framework and benchmark for wound classification, segmentation and localization that enables systematic evaluation of multi-task wound analysis under a common protocol. WILLIE harmonizes three public wound datasets into a shared benchmark and compares unified models across three scaling configurations against 10 single-task baselines. The best model achieves 91.88% classification accuracy, 91.41% Dice, and 96.23% AP@0.5 while producing all three outputs in a single forward pass. Beyond aggregate performance, our results show that segmentation-derived localization outperforms dedicated detection baselines in this benchmark, suggesting that box-based localization may be unnecessary for spatially coherent wound targets. Our findings highlight that effective multi-task learning in healthcare imaging depends not only on shared representations, but also on task formulation, compatibility, and benchmark design.

Figures & tables

Figure 1: Clinical wound assessment involves three coupled image analysis tasks: anatomical localization, boundary segmentation, and wound type classification. WILLIE provides a unified framework for producing these outputs with a single pipeline.

Figure 2: Construction of the WILLIE benchmark from three public wound datasets. FUSeg provides binary masks for segmentation and derived localization, AZH provides expert-annotated wound categories, and Medetec supplements underrepresented wound classes for classification. These sources are harmonized into a unified 5-class taxonomy and shared benchmark supporting classification, segmentation, and localization under a common evaluation protocol.

	Train			Test
Class	AZH	Medetec	Total	AZH	Weight
Diabetic	139	81	220	46	0.982
Pressure	100	170	270	34	0.802
Surgical	122	—	122	42	1.765
Venous	185	133	318	62	0.680
No Wound	150	—	150	50	1.434

Table 1: Classification class distribution with inverse-frequency training weights.

Figure 3: Overview of WILLIE architecture. A wound image is processed through shared pretrained backbones, cross-backbone fusion, multi-scale feature construction, and task-specific scale routing to produce three outputs: wound type classification, boundary segmentation, and localization.

Variant	Backbone(s)	Total	Trainable	Frozen	% Trainable
MINI	DINOv2-S	34.3M	12.3M	22.1M	35.7
BASE	DINOv2-L + ConvNeXt-L	520.4M	19.8M	500.6M	3.8
XL	DINOv2-L + ConvNeXt-L + SAM2-Hiera-L	762.5M	360.2M	402.3M	47.3

Table 2: Model variant parameter statistics.

Model	Type	Params	Acc (%)	F1 (%)	AUC
VGG-19	Baseline	124.9M	78.63	78.14	0.956
EfficientNet-B4	Baseline	18.5M	83.76	83.36	0.975
DINOv2+Linear	Ablation	22.2M	86.75	87.01	0.986
WILLIE-MINI (ens)	Unified	34.3M	88.90	87.60	0.976
ResNet-50	Baseline	24.6M	88.03	87.82	0.973
WILLIE-XL (TTA)	Unified	762.5M	91.88	90.73	0.986

Table 3: Classification comparison on the held-out AZH test set (

n=234

Model	Type	Params	Mean Dice	Median Dice
U-Net (Eff-B3)	Baseline	13.2M	84.59	91.20
WILLIE-MINI	Unified	34.3M	81.80	87.50
WILLIE-BASE	Unified	520.4M	86.36	92.89
MedSAM (ViT-B)	Baseline	93.7M	90.93	93.33
WILLIE-XL	Unified	762.5M	91.41	93.50
SAM2.1 (Hiera-S)	Baseline	46.1M	91.74	93.76

Table 4: Segmentation comparison on FUSeg validation (

n=400

s $\rightarrow$ l: segmentation-to-localization pipeline.
Model	Type	Params	AP@0.5	Prec.	Recall	F1
RT-DETR-L	Dedicated	32.8M	87.95	88.15	89.38	88.76
YOLOv8m	Dedicated	25.9M	91.22	87.45	88.47	87.96
WILLIE-MINI s $\to$ l	Unified	34.3M	84.60	89.03	88.66	88.84
WILLIE-BASE s $\to$ l	Unified	520.4M	89.91	92.40	95.40	93.87
WILLIE-XL s $\to$ l	Unified	762.5M	96.23	96.23	94.64	95.43

Table 5: Localization comparison on FUSeg validation (

n=400

Figure 4: Task-dependent scaling behavior across WILLIE variants. Classification accuracy improves from MINI to BASE and then saturates at XL, whereas segmentation Dice and localization AP@0.5 continue to improve with model scale.

Class	Prec.	Recall	F1	Support	Train
Diabetic	97.3	78.3	86.7	46	220
Pressure	76.5	76.5	76.5	34	270
Surgical	88.9	95.2	92.0	42	122
Venous	91.0	98.4	94.6	62	318
No Wound	98.0	100.0	99.0	50	150

Table 6: Per-class results for WILLIE-BASE (top-3 TTA).

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	MINI	BASE	XL
Full	86.6	88.2	89.8
$-$ F 2 DCA	–	87.8	89.9
$-$ WA-CSA	86.6	86.9	90.2
$-$ MoE	86.7	87.6	89.8
$-$ WTCS	86.1	87.0	89.5

Appendix

Table 7: Components ablation: classification accuracy (%). 5-fold CV, validation folds, no TTA.

Configuration	MINI	BASE	XL
Full	82.3	84.1	83.2
$-$ F 2 DCA	–	84.9	83.6
$-$ WA-CSA	82.9	84.7	84.0
$-$ MoE	83.5	83.8	82.1
$-$ WTCS	83.6	84.9	84.0

Appendix

Table 8: Component ablation: segmentation Dice (%). 5-fold CV, validation folds, no TTA.

Configuration	MINI	BASE	XL
Full	84.1	85.6	85.0
$-$ F 2 DCA	–	86.8	85.7
$-$ WA-CSA	85.5	86.7	85.2
$-$ MoE	85.3	85.6	85.0
$-$ WTCS	86.1	87.6	85.0

Appendix

Table 9: Component ablation: localization AP@0.5 (%). 5-fold CV, validation folds, no TTA.

Explore similar work

Automated multi-class wound assessment using dedicated instance segmentation models for boundary detection and classification
Date pendingMehedi Hasan Tusar, Fateme Fayyazbakhsh, Igor Melnychuk +1Lesion Segmentation Amodal Segmentation
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
WoundFormer: Multi-Scale Spatial Feature Fusion for Multi-Class Wound Tissue Segmentation
May 19, 2026Muhammad Ashad Kabir, Rabin DulalLesion Segmentation
School of Computing, Mathematics and Engineering, Charles Sturt University, NSW, Australia
Evaluation of Medical Vision Language Models HuluMed and MedGemma, and general purpose chatbots Gemma 3, ChatGPT Plus, and Claude Pro on real previously unseen wound images
Jun 16, 2026Yunzhe Xue, Mohammed Saim Ahmed Quadri, Neal Panse +2Medical Vision-Language Models Multimodal Clinical Data
Department of Data Science, New Jersey Institute of Technology, Newark, NJ, USA · Department of Computer Science, New Jersey Institute of Technology, Newark, NJ, USA · Vascular and Endovascular Surgery, Robert Wood Johnson Hospital, New Brunswick, NJ, USA

Date pending, cs.CVJ/K move · Enter open · S save

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

Mehedi Hasan Tusar, Fateme Fayyazbakhsh, Igor Melnychuk, Ming C. Leu

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

Lesion Segmentation Amodal Segmentation Yolo26 Delineation

Open paper AskPDF arXiv

Accurate wound classification (WC) and boundary segmentation are essential for guiding clinical decisions in chronic and acute wound management. However, most existing artificial intelligence (AI) models are limited, focusing on a narrow set of wound types, limited variations in wound severity, or a single task (segmentation or classification), which reduces their clinical applicability. This study presents two dedicated instance segmentation models based on You Only Look Once (YOLO)v11 that perform wound boundary segmentation (WBS) and WC across five clinically relevant wound types: burn injury (BI), pressure injury, diabetic foot ulcer, vascular ulcer, and surgical wound. A wound-type balanced dataset of 2,963 annotated images was created to train the models for both tasks, using five-fold cross-validation. Models trained on the original, non-augmented dataset performed consistently across folds, though BI detection accuracy was relatively low; augmenting the dataset with rotation, flipping, and variations in brightness, saturation, and exposure significantly improved performance, particularly for visually subtle BI cases. Among the tested variants, YOLOv11x achieved the best WBS performance (F1-score: 0.9341; mAP50: 0.9629). For WC, YOLOv11m achieved the highest mAP50 (0.9194) and mAP50-95 (0.6950), whereas YOLOv11l achieved the highest F1-score (0.8797). The lightweight YOLOv11n provided comparable accuracy at lower computational cost, making it suitable for resource-constrained deployments. Supported by confusion matrices and visual detection outputs, the results confirm robustness against complex backgrounds and high intra-class variability, demonstrating the potential of YOLOv11-based architectures for accurate, real-time wound analysis in clinical and remote care settings.

Abstract

Explore similar work

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

WoundFormer: Multi-Scale Spatial Feature Fusion for Multi-Class Wound Tissue Segmentation

Evaluation of Medical Vision Language Models HuluMed and MedGemma, and general purpose chatbots Gemma 3, ChatGPT Plus, and Claude Pro on real previously unseen wound images