Leveraging Deep Neural Networks for Automatic and Standardised

Wound Image Acquisition

Ana Filipa Sampaio

1 a

, Pedro Alves

1 b

, Nuno Cardoso

1 c

, Paulo Alves

2, 3 d

, Raquel Marques

3 e

Pedro Salgado

and Maria Jo

ao M. Vasconcelos

1 f

Fraunhofer Portugal AICOS, Porto, Portugal

School of Nursing Department, Universidade Cat

olica Portuguesa (UCP), Porto, Portugal

Centre for Interdisciplinary Research in Health, Instituto de Ci

encias da Sa

ude, UCP, Porto, Portugal

F3M Information Systems, S.A, Portugal

Keywords:

Skin Wounds, Mobile Health, Object Detection, Deep Learning, Mobile Devices.

Abstract:

Wound monitoring is a time-consuming and error-prone activity performed daily by healthcare professionals.

Capturing wound images is crucial in the current clinical practice, though image inadequacy can undermine

further assessments. To provide sufﬁcient information for wound analysis, the images should also contain

a minimal periwound area. This work proposes an automatic wound image acquisition methodology that

exploits deep learning models to guarantee compliance with the mentioned adequacy requirements, using a

marker as a metric reference. A RetinaNet model detects the wound and marker regions, further analysed by

a post-processing module that validates if both structures are present and veriﬁes that a periwound radius of

4 centimetres is included. This pipeline was integrated into a mobile application that processes the camera

frames and automatically acquires the image once the adequacy requirements are met. The detection model

achieved mAP@.75IOU values of 0.39 and 0.95 for wound and marker detection, exhibiting a robust detection

performance for varying acquisition conditions. Mobile tests demonstrated that the application is responsive,

requiring 1.4 seconds on average to acquire an image. The robustness of this solution for real-time smartphone-

based usage evidences its capability to standardise the acquisition of adequate wound images, providing a

powerful tool for healthcare professionals.

1 INTRODUCTION

The prevalence of chronic wounds is high world-

wide, being considered a major public health prob-

lem. Chronic wounds are those that do not progress

through a normal and timely sequence of repair

(Werdin et al., 2009) and can include pressure, ve-

nous, arterial, diabetic ulcers, and others. The study

of (Martinengo et al., 2019) indicates a pooled preva-

lence of 2.21 per 1000 population for chronic wounds

of mixed etiologies, and the prevalence for chronic

leg ulcers was estimated at 1.51 per 1000 population.

In the United States, chronic wounds affect the qual-

https://orcid.org/0000-0003-1770-4429

https://orcid.org/0000-0002-0372-4755

https://orcid.org/0000-0001-5736-7995

https://orcid.org/0000-0002-6348-3316

https://orcid.org/0000-0002-6701-3530

https://orcid.org/0000-0002-0634-7852

ity of life of nearly 2.5% of the total population, hav-

ing a signiﬁcant impact on healthcare (Sen, 2021). In

Portugal, the prevalence is 0.84 per 1000 population,

and patients with more than 80 years of age present a

prevalence of 5.86 (Passadouro et al., 2016).

Wound healing is a complex process that involves

several phases and proper monitoring of its evolution.

The rise of the digital health sector can help health-

care professionals properly monitor wound evolution

by storing clinical information in each monitorization,

including wound details and images. According to the

study of (Chen et al., 2020) current evidence suggests

that telemedicine has similar efﬁcacy and safety with

conventional standard care of chronic wounds.

An essential step for proper wound documenta-

tion is wound image capturing (Mukherjee et al.,

2017),(Anisuzzaman et al., 2022b); however, poor

image quality can negatively inﬂuence later assess-

ments. Many of the approaches reported in the lit-

erature for automatised wound monitoring address

Sampaio, A., Alves, P., Cardoso, N., Alves, P., Marques, R., Salgado, P. and Vasconcelos, M.

Leveraging Deep Neural Networks for Automatic and Standardised Wound Image Acquisition.

DOI: 10.5220/0012031200003476

In Proceedings of the 9th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2023), pages 253-261

ISBN: 978-989-758-645-3; ISSN: 2184-4984

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

253

the inadequate image quality using computer vision

methodologies to remove acquisition artefacts and

correct colour and perspective distortions in the cap-

tured images (Lu et al., 2017; Cui et al., 2019). Still,

the application of pre-processing strategies may be

insufﬁcient for some tasks since it does not ensure

that all the relevant structures needed for an adequate

wound assessment (e.g. the periwound skin) are prop-

erly visible. Subsequently, efforts to enhance the im-

age quality at the acquisition step have been con-

ducted by either providing acquisition guidelines or

using semi-automated acquisition processes.

Many of the attempts to automatise the acquisi-

tion of wound images resort to object detection mod-

els, namely convolutional neural networks (CNNs)

to locate the wound bed in the captured image ﬁeld.

The work in (Goyal et al., 2018) ﬁrst presented an

algorithm to detect and locate diabetic foot ulcers

in real-time. The authors trained a Faster R-CNN

model with Inception-V2, obtaining a mean average

precision (mAP) of 91.8%. In (Faria et al., 2020), a

methodology for automated image acquisition of skin

wounds via mobile devices is presented. The applica-

tion of image focus validation combined with a detec-

tion deep neural network (DNN) for wound bed local-

isation (an SSDLite model with a MobileNetV2 back-

bone) achieved an mAP at a 0.50 intersection over

union (IOU) threshold (mAP@.50IOU) of 86.46%

and an inference speed of 119 milliseconds (ms)

for smartphone usage, ensuring simultaneous image

quality and adequacy in real-time. More recently, a

similar wound localiser for mobile usage was pre-

sented in (Anisuzzaman et al., 2022a). Two detection

DNNs, SSD with a VGG16 backbone and YOLOv3,

were trained in a private dataset of 1010 images of

venous, pressure and diabetic foot ulcers, achieving

mAP values of 86.4% and 93.9%, respectively. In

(Scebba et al., 2022), the authors trained a RetinaNet

model with three different feature extractors, all pre-

trained on the Common Objects in Context (COCO)

dataset. The best performance was achieved with the

MobileNet backbone, leading to a test mAP@.50IOU

of 0.762.

Despite relying mainly on private datasets for the

models’ development, most of these approaches also

resorted to a public database, Medetec (Medetec,

2014), to expand the training set (Faria et al., 2020)

and to assess the performance of their approach in

other datasets (Anisuzzaman et al., 2022a). How-

ever, even being used in many of the aforementioned

works, the Medetec database does not provide infor-

mation regarding the wound region location in its im-

ages (its usage was enabled by private annotations).

The high detection performances reported by

these approaches highlight the potential of DNNs for

the analysis of wound images. Yet, the majority of

these works only use it to narrow the wound region for

further processing by other algorithms, such as seg-

mentation or classiﬁcation models. The exception is

the systems described in (Anisuzzaman et al., 2022a;

Faria et al., 2020), which integrate the trained wound

detectors in mobile wound acquisition applications.

Still, their frameworks do not use detection informa-

tion to analyse the adequacy of the captured images.

Thus, considering the importance of adequate

wound images and the lack of solutions that incorpo-

rate this validation in the image acquisition step, this

work presents an approach to standardise wound im-

age acquisition by automating the process while guar-

anteeing wound detection and periwound skin. For

this purpose, a new dataset of wound images with

reference markers was constructed and used to de-

velop a wound and marker detection model based on

a RetinaNet architecture with a MobileNetV2 back-

bone. This model was used as the basis for an image

adequacy validation module, responsible for ensur-

ing that the acquired images are compliant with the

requirements recommended by clinical experts: the

presence of the wound and marker regions and inclu-

sion of 4 centimetres (cm) of periwound skin in the

captured image. A mobile application incorporating

the proposed pipeline was implemented and tested on

several images, enabling a real-time, automated and

smooth process for image acquisition and assessment.

Therefore, this work advances the previous state-

of-the-art approaches by providing a detection model

for simultaneous wound and reference marker local-

isation, taking advantage of its potentialities to pro-

pose a simple yet effective strategy for validating

the adequacy of wound images, and integrating this

pipeline in a user-friendly mobile solution that aims

to facilitate monitoring procedures while ensuring the

quality of the wound images acquired.

2 METHODOLOGY

The proposed approach to automate the wound im-

age acquisition relies on the application of deep learn-

ing detection models for the correct localisation of

the wound bed and the reference marker in the ﬁeld

of view, in association with a domain-adapted post-

processing procedure to verify the presence of all the

relevant structures - open wound, marker and peri-

wound skin. This strategy ensures that all the in-

formation needed for a complete visual assessment

of the wound status and the automated extraction of

wound properties from the image are present in the

ICT4AWE 2023 - 9th International Conference on Information and Communication Technologies for Ageing Well and e-Health

254

captured images, providing a useful tool for image

acquisition both for manual clinical monitorisation of

wounds and for automated assessment solutions.

For the development of this methodology, a novel

dataset - the Clinical Wound Support (CWS) dataset

- was constructed, as described in section 2.1. The

succeeding sections detail the hyperparameter tuning

process used to ﬁnd the most promising model (sec-

tion 2.2.1), as well as the dataset preparation for the

training of the detection models (section 2.2.2), the

implementation of the adequacy validation module

(section 2.3) and the integration of the whole acqui-

sition pipeline in a mobile application (section 2.4).

2.1 Dataset Construction

The scarce datasets with reference markers available

in the literature are based on complex and expensive

markers and do not provide annotations regarding the

marker regions (Yang et al., 2016). Therefore, a novel

dataset comprising images of wounds and reference

markers was constructed. The images were acquired

by clinical professionals from nine Portuguese health-

care institutions using a smartphone application de-

veloped for this purpose using commercially available

adhesive reference markers

. These reference mark-

ers were selected due to their ready accessibility, ease

of usage and standardised size (2 × 2 cm).

The dataset comprises monitorisations of differ-

ent types of chronic wounds: diabetic foot ulcers,

pressure ulcers and leg ulcers (venous, arterial and

mixed), among others. Although that was not always

possible, in most cases, the wound was followed up

weekly over a 4-week period, resulting in a maximum

of 5 monitorisations per wound to capture varying

healing states of each type of wound. The resulting

images were annotated in terms of wound and marker

boundaries by three specialists in the ﬁeld of wound

care, each responsible for a distinct subset of wound

images. The dataset distribution is presented in ta-

ble 1, and annotated examples are depicted in ﬁg. 1.

Table 1: Composition of the acquired CWS dataset, in terms

of wounds, monitorisations and captured images.

Wound

type

Pressure

ulcer

Venous

leg ulcer

Arterial

leg ulcer

Leg ulcer of

unknown

etiology

Diabetic

foot ulcer

Other

ulcer

types

Total

No. wounds 29 10 2 4 5 1 51

No. monitor. 61 36 7 18 17 6 145

No. images 65 52 8 20 17 6 168

2.2 Model Development

The wound and marker localisation module was de-

veloped using a RetinaNet detection architecture with

https://www.apli.com/en/product/18326

Figure 1: Illustrative examples of skin wound images and

respective ground truth localisation (open wound in cyan

and reference marker in green).

a MobileNetV2 backbone network. MobileNetV2

(Sandler et al., 2018) belongs to a family of CNNs

optimised for mobile deployment through the appli-

cation of lightweight ﬁltering operations and an in-

verted residual structure, being able to provide com-

petitive performances with a reduced number of pa-

rameters and a lower computational complexity. Reti-

naNet is a one-stage detector that combines off-the-

shelf CNNs with a feature pyramid network to enable

the detection of objects at multiple scales and uses a

focal loss function to account for possible data imbal-

ances during training (Lin et al., 2018). The Tensor-

Flow Python implementation of the model was used,

resorting to the Object Detection API for model train-

ing and optimisation. This model was selected con-

sidering the need for deployment in a mobile scenario,

as it offers a good trade-off between detection per-

formance and inference time while being easily ex-

ported to a smartphone-compatible format (Tensor-

Flow Lite

). This was assessed based on the bench-

marks in (Huang et al., 2017; Ten, 2021).

2.2.1 Model Optimisation

The detection model conﬁguration was optimised us-

ing a random search hyperparameter tuning process,

applied instead of grid search optimisation due to time

constraints. Six different batch sizes and learning rate

combinations were analysed. These two hyperparam-

eters were selected for tuning due to their high im-

pact on the training results. The corresponding values

were randomly sampled from adequate ranges based

on the recommendations of (Bengio, 2012). Batch

sizes of 8 and 16 were considered, as these provide

sufﬁcient training instances in each learning itera-

tion while complying with the dynamic memory con-

straints of the hardware used. The learning rate val-

ues were sampled from a logarithmically scaled range

between 10

−6

and 0.01; a constant learning rate was

considered for all the iterations of each experiment.

Also, in view of the impact of the input image size on

the computational demands of the detection models

and the other processing steps, two image sizes were

assessed: 320 × 320 and 512 × 512 pixels. These hy-

perparameter combinations were tested for both input

https://www.tensorﬂow.org/lite

Leveraging Deep Neural Networks for Automatic and Standardised Wound Image Acquisition

255

Table 2: Hyperparameter values tested for the six experi-

ments conducted to optimise the detection model.

Experiment Batch size Learning rate Image size

1 8 7.56463E-03 320 / 512

2 8 1.12332E-05 320 / 512

3 16 5.59081E-04 320 / 512

4 8 6.57933E-05 320 / 512

5 8 1.14976E-04 320 / 512

6 16 7.56463E-03 320 / 512

sizes, resulting in the experiments shown in table 2.

A 3-fold cross-validation was followed for the hy-

perparameter tuning experiments, with details regard-

ing the data preparation being provided in the next

section. The evaluation of the tuning results was con-

ducted on the validation subsets of each fold, consid-

ering performance metrics appropriate for object de-

tection tasks - mAP and average recall (AR), com-

puted for IOU thresholds of 0.50 and 0.75 in the

case of mAP, and for a maximum number of 10 de-

tections for AR. A higher focus was attributed to

mAP@.75IOU because the higher IOU threshold im-

posed leads to a more representative metric of the

model’s ability to provide accurate regions of inter-

est; mAP@.50IOU was evaluated for the state-of-the-

art comparison (Goyal et al., 2018; Faria et al., 2020;

Anisuzzaman et al., 2022a; Scebba et al., 2022).

Each model was optimised by applying the Adam

optimisation algorithm, using the Huber regression

and focal cross-entropy functions for the localiza-

tion and classiﬁcation losses, respectively. Although

each model was trained for 5000 epochs, during this

process, its validation performance was periodically

monitored, and the network parameters that yielded

the optimal validation mAP@.75IOU were consid-

ered as ﬁnal for the respective experiment to pre-

vent overﬁtting. The network was initialised using

the parameters resulting from pre-training in the pub-

lic COCO dataset (Lin et al., 2014). All the non-

mentioned training conditions and hyperparameters

were set to default values provided with the model

implementation, as speciﬁed in (Con, 2022).

2.2.2 Data Preparation

To develop the model, 75% of the dataset was used

for training, with the remaining 25% being separated

as test data for the ﬁnal model assessment. For the

hyperparameter tuning experiments, the training data

was split into three distinct folds of train and valida-

tion sets. The division of the data instances in the sev-

eral subsets was performed in a stratiﬁed manner, at

the level of the wounds and not the images, to ensure

that all images of each wound were kept in the same

subset and that the training and test subsets comprise

representative examples of each wound type, body

location and skin phototype. This process resulted

in a distribution of 98 training images from 29 dif-

ferent wounds and 70 test images from 22 wounds.

The images, already square as a result of their ac-

quisition process, were resized according to the input

size needed for the trained network. Geometric image

transformations, in particular 90

◦

, 180

◦

and 270

◦

ro-

tations, and vertical and horizontal ﬂips, were applied

dynamically during training to augment the instances.

2.2.3 Final Model Training and Mobile

Deployment

The ﬁnal model conﬁguration was identiﬁed as the

one that resulted in the highest performance met-

rics. After the determination of the best experience

from hyperparameter tuning, the detection model was

re-trained using the whole training data considering

the corresponding hyperparameter values, for a ﬁxed

number of epochs. The number of training epochs

was established as the average number of epochs

needed for the convergence of the models trained for

each cross-validation fold of the selected experiment.

The trained model parameters were then quantised

and optimised for mobile deployment using the Ten-

sorﬂow Lite framework. The best model of each input

size was deployed in a mobile device and compared

to assess its inﬂuence on the model’s inference speed

and its impact on the mobile acquisition workﬂow.

2.3 Image Adequacy Validation

The validation of the image adequacy was performed

based on the outputs of the detection model by as-

sessing if both the open wound and reference marker

were detected. In addition to the open wound, the

surrounding area, known as the periwound area, also

contains important information regarding the state of

the wound. Therefore, there is a need to ensure that

the captured image also contains this area. Following

experts’ recommendation, 4 cm in all directions of the

open wound border were considered.

Given the marker size of 2× 2 cm, the largest side

of the detected marker’s bounding box was used as

a reference, ensuring that, even if the marker is not

perpendicular to the camera, the side closer to 2 cm

will be used. Then, the veriﬁcation is performed by

ensuring that there is a distance of at least 2 times

the reference is present around the four sides of the

wound’s bounding box.

ICT4AWE 2023 - 9th International Conference on Information and Communication Technologies for Ageing Well and e-Health

256

2.4 Mobile Application

The methodology presented in the previous subsec-

tions was deployed as an Android application running

on a smartphone with the aim of supporting health-

care professionals in wound monitoring procedures.

This application allows the automatic acquisition

of skin wound images in an easy and intuitive manner

while providing real-time feedback about the wound

bed and marker’s localisation and guaranteeing the

proper periwound area presence (see ﬁg. 2).

Figure 2: Automatic wound image acquisition diagram.

The user is ﬁrst guided to centre the skin wound

in the camera feed through a central target while the

wound and marker localisation module is running

(ﬁg. 3 left). In the detection module, the raw im-

ages are cropped to a square at its centre with a height

equal to its total width and resized to the model’s in-

put size (pre-processing, see ﬁg. 2). Once the de-

tection is complete, the user receives the indication

through the left icon of the screen, and the adequacy

validation module (post-processing block in ﬁg. 2)

starts running (ﬁg. 3 right). The detection and ad-

equacy validation modules keep running frame by

frame until either one frame fails the detection (there-

fore restarting the process) or the acquisition is suc-

cessfully completed. An image is automatically ac-

quired when 20 consecutive frames or all the frames

acquired in the span of 1250ms detect a wound and

a marker with a score of at least 50% each. In case

the developed modules fail, the user is always capa-

ble of manually acquiring an image by clicking on the

bottom right button on the screen. After the acquisi-

tion, to make sure the user did not move after the last

frame was processed, the acquired image is also veri-

ﬁed with the detection and adequacy validation mod-

ules. In case of failure, the user has the option to retry

the acquisition or to still continue with the previously

acquired image.

3 RESULTS AND DISCUSSION

This section presents the results of the model optimi-

sation experiments, as well as an in-depth analysis of

the best-performing model in terms of detection per-

formance and mobile inference metrics.

Figure 3: Application screenshots of real-time wound local-

ization (left) and image adequacy validation (right).

3.1 Hyperparameter Tuning

To identify the most adequate hyperparameter com-

bination, the performance achieved in the differ-

ent experiments conducted for each image size was

compared in terms of mAP@.75IOU and AR@10,

for each detected class (open wound or reference

marker) and total. The average cross-validation

mAP@.75IOU of the hyperparameter tuning experi-

ments is presented in ﬁg. 4. Although the AR@10

results were also taken into account in the com-

parative analysis of the six experiments, those val-

ues exhibited a similar tendency as the ones of

mAP@.75IOU and are not graphically represented.

Figure 4 shows that all experiments resulted in a sim-

ilar performance for the reference marker detection,

so the best hyperparameter combination was selected

based mostly on the wound detection ability provided.

Thus, through the analysis of the values reported for

the wound class, experiment 5 yielded the highest

mAP@.75IOU and AR@10 metrics for both image

sizes, so the associated hyperparameters were consid-

ered for the ﬁnal detection model implementation.

For this training setup, the model trained on

the 320 × 320 images obtained cross-validation

mAP@.75IOU and AR@10 values of 0.46699 ±

0.01170 and 0.48186 ± 0.03115 for the wound class,

and 0.99010 ± 0.01400 and 0.83523 ± 0.00923 re-

garding the detection of the marker. Similarly,

the detection performance of the 512 × 512 model

resulted in mAP@.75IOU and AR@10 metrics of

0.46107 ± 0.04047 and 0.51709 ± 0.06306 for the

open wounds, and 0.99928 ± 0.00102 and 0.86706 ±

0.00871 for the marker class. Concerning the

cross-validation mAP@.50IOU for the open wound

class, the 320 × 320 network achieved an average of

0.66188 ± 0.05608, while the 512 × 512 model ob-

tained 0.78117 ± 0.11202; the mAP@.50IOU values

for the marker class are approximately equal to the

Leveraging Deep Neural Networks for Automatic and Standardised Wound Image Acquisition

257

Figure 4: Average class-wise 3-fold cross-validation mAP@.75IOU results obtained with the different hyperparameter set-

tings tested in each experiment for the two image sizes analysed. The represented values were obtained as the average

validation results for the three cross-validation splits. The total mAP was determined by averaging the results of the two

classes (wound and marker). The error bars correspond to the standard deviation over all splits.

mAP@.75IOU values reported. Both models exhibit

superior performance for marker detection in relation

to the open wound class, possibly due to the simplic-

ity of the marker shape and the lower variability of

the marker’s regions in the dataset. Moreover, de-

spite providing good performance for the identiﬁca-

tion of wounds in the images (materialised in the high

mAP@.50IOU values achieved), the models demon-

strated an impaired ability to accurately distinguish

the region of interest corresponding to the whole open

wound, as indicated by the lower mAP obtained when

imposing a stricter IOU threshold of 0.75.

3.2 Selection of the Best Model

Table 3 shows the test results of the ﬁnal RetinaNet

detection model for each class after being re-trained

on the whole training set with the optimal hyperpa-

rameter set-up (batch size of 8 and learning rate of

1.14976 × 10

−4

). Comparing the test metrics with

the cross-validation results presented in the previ-

ous section, we can observe that there was a general

decline in the models’ performance, both in terms

of detection accuracy and sensitivity (reﬂected by

the mAP@.50IOU and AR@10 metrics, correspond-

ingly), even though this deterioration was less se-

vere in the case of the marker class. This difference

may be due to the occurrence of overﬁtting during

model training or to the variability of properties in

the training and test set images. Moreover, the dis-

crepancy between the cross-validation and test perfor-

mances was more expressive for the 320 × 320 model.

Therefore, despite the similar cross-validation results

demonstrated by the models trained with the two im-

age sizes, in the test set the 512 × 512 demonstrated

considerably higher detection robustness.

To assess the applicability of each detection model

for real-time usage, both models were deployed on

Table 3: Class-wise detection performance of the best mod-

els, re-trained in the overall training set and evaluated using

the test data from the CWS dataset.

Evaluated metric/ Image size 320 × 320 512 × 512

mAP@.75IOU

Wound 0.23712 0.39141

Marker 0.92079 0.95050

AR@10

Wound 0.36563 0.47031

Marker 0.79412 0.83971

Table 4: Processing times (in milliseconds) of the differ-

ent steps of the image acquisition module using the best-

performing models associated with each input image size,

assessed in mid and high-range (MR, HR) smartphones.

320 × 320 512 × 512

MR HR MR HR

Total time per frame 132.00 103.43 393.52 274.79

Pre-processing 1.21 1.28 1.39 1.47

Detection inference 130.26 101.53 391.41 272.76

Adequacy validation 0.00 0.00 0.06 0.02

Total acquisition time 1313.98 1296.13 1400.63 1394.67

Avg. no. frames per acq. 12 12 5 6

Model size 10.9 MB 11.2 MB

two mobile devices (mid and high range), namely Xi-

aomi Poco X3 NFC and Samsung S21. Since the ac-

quisition of each image involves the analysis of sev-

eral camera frames by the application, the time re-

quired by each processing step - the pre-processing

to get compatible image shape and dimensions, the

detection model’s inference and the validation of the

image structures’ adequacy (open wound, reference

marker and periwound area) - was examined for all

the frames analysed, along with the total acquisition

time and the average number of frames processed for

each image acquired. The results of the mobile as-

sessment are presented in table 4. This evaluation was

carried out by using the developed image acquisition

application to capture images of the test set images.

From the times presented for each step, it is pos-

sible to verify that the inference time of the detection

ICT4AWE 2023 - 9th International Conference on Information and Communication Technologies for Ageing Well and e-Health

258

model is the most time-consuming step of the acqui-

sition process, being the major contributor to the time

taken to analyse each frame and, subsequently, for the

whole image acquisition ﬂow. Despite the higher in-

ference times required by the 512 × 512 model (in re-

lation not only to the 320 × 320 model but also to the

approach described in (Faria et al., 2020)), it is still

able to provide a responsive and smooth acquisition

process in both high and mid-range devices, with a

sufﬁcient amount of evaluated frames in the stipulated

acquisition time. This, in combination with its supe-

rior detection performance in the test set, motivated

its selection as the ﬁnal model to be deployed in the

ﬁnal acquisition application.

3.3 Final Model Assessment

To allow the comparison with the detection ap-

proaches described in the state of the art and men-

tioned in section 1, the mAP@.50IOU metric com-

puted for the wound class was also analysed. Compar-

ing the cross-validation mAP@.50IOU of the devel-

oped model (0.78117±0.11202) with the correspond-

ing performances reported in (Goyal et al., 2018;

Faria et al., 2020) (0.918 and 0.865), it is possible to

verify that, despite the lower metric value achieved,

it is associated with a similar magnitude order. In

contrast, the obtained test metric (0.66829) exhibits

a greater deviation in relation to the test performances

of (Anisuzzaman et al., 2022a; Scebba et al., 2022)

(0.939 and 0.762), which might be due to the pres-

ence of borderline wound cases in the test set, further

discussed below. Still, the direct comparison of these

metrics does not provide a fair analysis of the differ-

ent models’ performance since they were evaluated in

different datasets, whose information is not available.

Moreover, the models proposed in other works were

solely focused on the detection of the open wound

area, while this work provides a model that is able

to detect directly both the open wound and reference

marker. Hence, future experiments should be con-

ducted to enable a more even comparison, by training

or at least evaluating the models in the same dataset.

In addition, to validate the detection ability of

the network selected as the ﬁnal model, its predic-

tion results in the test images were analysed in detail.

Figure 5 comprises instances for which the detection

of the model was not as satisfactory, whereas ﬁg. 6

presents examples of correct detections.

The inspection of the test predictions shows that

the model’s detection ability is ﬂawed under very

speciﬁc acquisition conditions, providing insights re-

garding the possible causes of the inferior test perfor-

mance displayed. Particularly, it demonstrated difﬁ-

(a) (b) (c)

(d) (e) (f)

Figure 5: Examples of test images that highlight the perfor-

mance limitations of the 512 × 512 model. These include

images acquired in a skewed perspective (a and b), superﬁ-

cial wounds (c and d) and wounds in the open/healed tran-

sition (e and f). The ground truth annotations are marked

in dark blue, while light blue and green boxes outline the

wound and marker regions detected by the network.

culty in correctly detecting the marker and/or wound

region when the image is captured from a skewed per-

spective, a common occurrence in diabetic foot ulcer

cases, as showed in ﬁgs. 5a and 5b. In some cases

of more superﬁcial leg ulcers, especially with hetero-

geneous lighting conditions and the presence of re-

ﬂections, the model is not capable of detecting the

wound region (ﬁg. 5c), or provides only a partial de-

tection that does not include the whole region of in-

terest (ﬁg. 5d), which might affect the veriﬁcation of

the periwound area in the photo and be an obstacle to

the usage of this information to further extract wound

properties in automated systems. Notwithstanding,

these limitations can be overcome in the acquisition

moment by using adequate lighting and adjusting the

smartphone’s positioning in relation to the wound.

The model also misdetected wounds that were re-

cently healed or close to healing, by incorrectly de-

tecting the scab region in ﬁg. 5f as an open wound

and by missing the detection of the small wound re-

gion in ﬁg. 5e. It is worth highlighting that these cases

are ambiguous even for trained clinical experts, so this

limitation is not critical for a practical scenario.

In spite of failing in some of the described circum-

stances, the model presented a satisfactory detection

performance in a varied set of conditions, being able

to detect both the wound and marker in images ac-

quired from a skewed perspective (ﬁg. 6c) and in-

consistent lighting (such as the shadows present in

ﬁg. 6d). It demonstrated a consistent detection ability

for the multiple wound types and body locations of

the dataset - observable in the detections for hip pres-

sure ulcers (ﬁgs. 6a, 6d and 6e), leg ulcers (ﬁg. 6b)

Leveraging Deep Neural Networks for Automatic and Standardised Wound Image Acquisition

259

(a) (b) (c)

(d) (e) (f)

Figure 6: Examples of test images with correct predictions

generated by the 512 × 512 model in varied settings. These

comprise images from different wound types and locations

(hip pressure ulcers in a, d and e, leg ulcer in b and di-

abetic foot ulcer in c), and images with healed wounds (f).

The ground truth annotations are marked in dark blue, while

light blue and green boxes outline the wound and marker re-

gions detected by the network.

and diabetic foot ulcers (ﬁg. 6c) -, while showing a

reduced number of false positives, as it is possible

to observe by the lack of wound area detections in

images without open wounds (ﬁg. 6f). Remarkably,

notwithstanding the low representation of higher pho-

totypes in the dataset, it also maintained a good detec-

tion ability for varying skin phototypes (displayed by

the heterogeneous types of skin in ﬁg. 6), evidencing

its robustness and suitability for usage in clinical sce-

narios.

Thus, regardless of the apparently superior perfor-

mances reported in the literature, the developed model

was able to locate the wound and reference marker

regions consistently for different image acquisition

conditions. Its combination with the image valida-

tion module and the integration of this pipeline in a

mobile-based solution make this the ﬁrst methodol-

ogy (to the authors’ knowledge) to exploit the poten-

tial of detection architectures to ensure wound image

adequacy in the acquisition process, providing a real-

time acquisition system that captures images with suf-

ﬁcient information for the analysis of wound healing

by human experts and for the development of auto-

mated wound characterisation algorithms.

4 CONCLUSIONS

This work presented a pipeline for the automation

of wound image acquisition and adequacy validation,

which leverages the potential of deep detection neural

networks to guarantee that relevant skin regions are

captured while providing a metric reference, both el-

ements of great utility to support further wound anal-

ysis tasks.

A deep neural network - RetinaNet with a Mo-

bileNetV2 backbone - was used to detect the open

wound region and a reference marker, to be used as

a metric reference. This model was developed using

a new dataset of wounds and markers of diverse types

and characteristics. Two models associated with dif-

ferent image sizes of 320 × 320 and 512 × 512 were

tuned and deployed in mobile devices. Their compari-

son showed that the 512 × 512 model offers a superior

detection performance while still providing inference

times suitable for real-time image acquisition, being

selected as the ﬁnal model. Although this model

yielded detection metrics slightly inferior to those re-

ported in the literature for similar models (evaluated

in other datasets), it demonstrated a satisfactory abil-

ity to detect wounds and markers in a variety of ac-

quisition conditions, evidencing its robustness.

The developed detection model was used as the

basis for an image adequacy validation module, re-

sponsible for ensuring the presence of the reference

marker, wound area and recommended 4 cm radius of

periwound skin (validated using the marker’s metric

reference) in the captured picture. A mobile appli-

cation incorporating both modules was implemented

and tested for the acquisition of several wound im-

ages. This assessment showed that the proposed ap-

proach contributed to the acquisition of adequate im-

ages through an easily usable interface that stream-

lines the acquisition process. Nonetheless, the peri-

wound radius veriﬁcation could be improved by get-

ting a more rigorous marker segmentation and using

the resulting sides or diagonal as a metric reference

instead of its bounding box, as well as resorting to

perspective correction methods to account for possi-

ble image distortions.

In spite of the identiﬁed reﬁnement possibilities,

the integration of the developed detection network

with the image adequacy validation module provides

a powerful tool to assist wound image acquisition, en-

abling its standardisation. Consequently, it facilitates

the acquisition of more adequate images, with the rep-

resentation of all the structures relevant for wound

healing assessment, offering an invaluable asset for

both the visual inspection by clinical experts and the

automated wound properties extraction in more so-

phisticated systems. Although its accuracy in the

wound region identiﬁcation should be improved for

the latter application scenario (through the expansion

of the dataset with private or public images or the ex-

ploration of new detection architectures), the robust-

ness demonstrated in varied acquisition settings, com-

ICT4AWE 2023 - 9th International Conference on Information and Communication Technologies for Ageing Well and e-Health

260

bined with its user-friendly, responsive mobile inter-

face, proved it is suitable for usage by clinical profes-

sionals in real-world scenarios.

ACKNOWLEDGEMENTS

This work was done under the scope of Clini-

calWoundSupport - ”Wound Analysis to Support

Clinical Decision” project (POCI-01-0247-FEDER-

048922), according to Portugal 2020 is co-funded

by the European Regional Development Fund from

European Union, framed in the COMPETE 2020.

The authors would like to thank Paula Teixeira from

Unidade Local de Sa

ude de Matosinhos for the col-

laboration in the data annotation process.

REFERENCES

(2021). TensorFlow 2 Detection Model Zoo. https://

github.com/tensorﬂow/models/blob/master/research/

object detection/g3doc/tf2 detection zoo.md.

(2022). Conﬁguration settings for the mod-

els of the tensorﬂow object detection API.

https://github.com/tensorﬂow/models/tree/master/

research/object detection/conﬁgs/tf2.

Anisuzzaman, D., Patel, Y., Niezgoda, J. A., Gopalakr-

ishnan, S., and Yu, Z. (2022a). A mobile app for

wound localization using deep learning. IEEE Access,

10:61398–61409.

Anisuzzaman, D., Wang, C., Rostami, B., Gopalakrishnan,

S., Niezgoda, J., and Yu, Z. (2022b). Image-based

artiﬁcial intelligence in wound assessment: A system-

atic review. Advances in Wound Care, 11(12):687–

709.

Bengio, Y. (2012). Practical recommendations for gradient-

based training of deep architectures. In Neural net-

works: Tricks of the trade, pages 437–478. Springer.

Chen, L., Cheng, L., Gao, W., Chen, D., Wang, C., Ran, X.,

et al. (2020). Telemedicine in chronic wound man-

agement: systematic review and meta-analysis. JMIR

mHealth and uHealth, 8(6):e15574.

Cui, C., Thurnhofer-Hemsi, K., Soroushmehr, R., Mishra,

A., Gryak, J., Dom

ınguez, E., Najarian, K., and

opez-Rubio, E. (2019). Diabetic wound segmenta-

tion using convolutional neural networks. In 2019

41st Annual International Conference of the IEEE En-

gineering in Medicine and Biology Society (EMBC),

pages 1002–1005. IEEE.

Faria, J., Almeida, J., Vasconcelos, M. J. M., and Rosado, L.

(2020). Automated mobile image acquisition of skin

wounds using real-time deep neural networks. In Med-

ical Image Understanding and Analysis: 23rd Confer-

ence, MIUA 2019, Liverpool, UK, July 24–26, 2019,

Proceedings 23, pages 61–73. Springer.

Goyal, M., Reeves, N. D., Rajbhandari, S., and Yap, M. H.

(2018). Robust methods for real-time diabetic foot

ulcer detection and localization on mobile devices.

IEEE journal of biomedical and health informatics,

23(4):1730–1741.

Huang, J., Rathod, V., Sun, C., Zhu, M., Korattikara, A.,

Fathi, A., Fischer, I., Wojna, Z., Song, Y., Guadar-

rama, S., et al. (2017). Speed/accuracy trade-offs for

modern convolutional object detectors. In Proceed-

ings of the IEEE conference on computer vision and

pattern recognition, pages 7310–7311.

Lin, T., Goyal, P., Girshick, R., He, K., and Dollar, P.

(2018). Focal loss for dense object detection. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 42(2):318–327.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Euro-

pean conference on computer vision, pages 740–755.

Lu, H., Li, B., Zhu, J., Li, Y., Li, Y., Xu, X., He, L., Li,

X., Li, J., and Serikawa, S. (2017). Wound intensity

correction and segmentation with convolutional neu-

ral networks. Concurrency and computation: practice

and experience, 29(6):e3927.

Martinengo, L., Olsson, M., Bajpai, R., Soljak, M., Upton,

Z., Schmidtchen, A., Car, J., and J

arbrink, K. (2019).

Prevalence of chronic wounds in the general popula-

tion: systematic review and meta-analysis of observa-

tional studies. Annals of epidemiology, 29:8–15.

Medetec (2014). Wound database. data re-

trieved from http://www.medetec.co.uk/ﬁles/

medetec-image-databases.html.

Mukherjee, R., Tewary, S., and Routray, A. (2017). Di-

agnostic and prognostic utility of non-invasive multi-

modal imaging in chronic wound monitoring: a sys-

tematic review. Journal of medical systems, 41:1–17.

Passadouro, R., Sousa, A., Santos, C., Costa, H., and

Craveiro, I. (2016). Characteristics and prevalence of

chronic wounds in primary health care. Journal of the

Portuguese Society of Dermatology and Venereology,

74(1):45–51.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2018). Mobilenetv2: Inverted residu-

als and linear bottlenecks. In Proceedings of the IEEE

conference on computer vision and pattern recogni-

tion, pages 4510–4520.

Scebba, G., Zhang, J., Catanzaro, S., Mihai, C., Dis-

tler, O., Berli, M., and Karlen, W. (2022). Detect-

and-segment: A deep learning approach to automate

wound image segmentation. Informatics in Medicine

Unlocked, 29:100884.

Sen, C. K. (2021). Human wound and its burden: updated

2020 compendium of estimates. advances in wound

care, 10(5):281–292.

Werdin, F., Tennenhaus, M., Schaller, H.-E., and Ren-

nekampff, H.-O. (2009). Evidence-based manage-

ment strategies for treatment of chronic wounds.

Eplasty, 9.

Yang, S., Park, J., Lee, H., Kim, S., Lee, B.-U., Chung,

K.-Y., and Oh, B. (2016). Sequential change of

wound calculated by image analysis using a color

patch method during a secondary intention healing.

PloS one, 11(9):e0163092.

Leveraging Deep Neural Networks for Automatic and Standardised Wound Image Acquisition

261