Unsupervised diabetic foot monitoring techniques

Abstract

A significant amount of research, involving computerized methods, has been initiated the last few years regarding the identification and prevention of Diabetes Foot Ulceration (DFU). In this paper, the spatial analysis of the raw data is investigated. The major expectations were the indication of regions of interest and the extraction of a more reliable understanding, regarding the captured information. Towards this direction, unsupervised learning approaches were used for image segmentation purposes. According to the experimental results, high-level features can be used to segment coarse images, grouping together areas with skin irregularities on patient's foot. In practice, there are (or can be calculated) appropriate features, over RGB images, that will facilitate the detection of problematic/high-risk regions on a foot. Yet, unsupervised approaches should not be considered as viable monitoring solutions both in terms of time and accuracy. However, the proposed approach could potentially be used to assist the detection process resulted by supervised Deep Learning techniques.

Full text

Unsupervised diabetic foot monitoring techniques
I. N. Tzortzis
National Technical University of
Athens
Greece
itzortzis@mail.ntua.gr
A. Davradou
National Technical University of
Athens
Greece
adavradou@gmail.com
E. Protopapadakis
National Technical University of
Athens
Greece
eftprot@mail.ntua.gr
M. Kaselimi
National Technical University of
Athens
Greece
mkaselimi@mail.ntua.gr
N. Doulamis
National Technical University of
Athens
Greece
adoulam@cs.ntua.gr
A. Aggeli
National and Kapodistrian University
of Athens
Greece
A. Lazaris
National and Kapodistrian University
of Athens
Greece
ABSTRACT
A signicant amount of research, involving computerized methods,
has been initiated the last few years regarding the identication
and prevention of Diabetes Foot Ulceration (DFU). In this paper, the
spatial analysis of the raw data is investigated. The major expecta-
tions were the indication of regions of interest and the extraction of
a more reliable understanding, regarding the captured information.
Towards this direction, unsupervised learning approaches were
used for image segmentation purposes. According to the experimen-
tal results, high-level features can be used to segment coarse images,
grouping together areas with skin irregularities on patient’s foot. In
practice, there are (or can be calculated) appropriate features, over
RGB images, that will facilitate the detection of problematic/high-
risk regions on a foot. Yet, unsupervised approaches should not be
considered as viable monitoring solutions both in terms of time
and accuracy. However, the proposed approach could potentially
be used to assist the detection process resulted by supervised Deep
Learning techniques.
CCS CONCEPTS
•Information systems →
Process control systems;
•Computing
methodologies →Image segmentation.
KEYWORDS
diabetic foot ulcer, neural networks, clustering, low-level features,
high-level features
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specic permission
and/or a fee. Request permissions from permissions@acm.org.
PETRA ’22, June 29-July 1, 2022, Corfu, Greece
©2022 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-9631-8/22/06. . . $15.00
https://doi.org/10.1145/3529190.3534723
ACM Reference Format:
I. N. Tzortzis, A. Davradou, E. Protopapadakis, M. Kaselimi, N. Doulamis,
A. Aggeli, and A. Lazaris. 2022. Unsupervised diabetic foot monitoring
techniques. In The15th International Conference on PErvasive Technologies
Related to Assistive Environments (PETRA ’22), June 29-July 1, 2022, Corfu,
Greece. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3529190.
3534723
1 INTRODUCTION
Self-monitoring can highly assist in preventing or postponing the
Diabetic Foot Monitoring (DFU), by capturing insights that can
lead to an early-diagnosis. Nowadays, the majority of these DFU in-
dication signs can be detected and, consequently, monitored, using
various optical sensors [
6
]. Partitioning an image into meaningful
parts is termed as Medical Image Segmentation (MIS). It is an impor-
tant step towards automating medical image analysis and is crucial
for a variety of medical imaging applications, such as computer
aided diagnosis, therapy planning and delivery, and computer aided
interventions. However, the existence of noise, low contrast and
object complexity in medical images are critical obstacles that stand
in the way of achieving an ideal segmentation system.
The main scope of this paper is to provide a set of signal pro-
cessing and Machine Learning (ML) tools to be applied in medical
images of diabetic foot to detect Regions of Interest (ROIs) and to
extract meaningful features to boost model performance in DFU
detection and monitoring. Towards this direction, a set of unsuper-
vised learning approaches have been considered, including Low-
Level (LL) feature extraction, clustering approaches, and stacked
visual auto-encoders.
The remainder of the paper is structured as follows: section 2 de-
scribes various approaches on area segmentation in medical images.
Section 3 provides details on the adopted approaches employing
LL features and image clustering approaches. Section 4 demon-
strates how High-Level (HL) features, using Deep Learning (DL)
approaches can be used for image segmentation purposes. Finally,
section 5 concludes this work.
1
608

PETRA ’22, June 29-July 1, 2022, Corfu, Greece Tzortzis, et al.
Filters
... ...
Labels
Label 1 Mask
Label 2 Mask
Label M Mask
Feature 1
Feature 2
Feature N
Draw bounding boxes
ROI contours detection
ROI contours detection
ROI contours detection
+
Multi-channelimage
Features Extraction
Clustering Application
Clustering
Bounding Boxes Creation
Clustered Img with Bounding Boxes
Original Image
Figure 1: The proposed methodology - Processing pipeline
2 RELATED WORK
Various segmentation algorithms have been proposed in the last
decade which can be broadly categorized into three groups: feature-
based, region-based, and graph-based methods [
11
]. Feature-based
algorithms generally use the color or texture information to group
the similar features into well separated clusters [
14
]. This separa-
tion issue is handled by a pre-dened distance measure. Clustering
based image segmentation approaches are generally ecient. How-
ever, some of them do not consider the spatial information which
causes under-segmentation. Region based segmentation methods
can preserve the edge or spatial information to produce more ho-
mogeneous and compact regions [
7
], [
1
]. The watershed algorithm
[
15
] is a popular region-based segmentation approach. However, it
might lead to over-segmentation. This drawback can be eliminated
by using a further merging procedure to construct more meaningful
regions [9]. Graph-based techniques are quite ecient in the eld
of image segmentation because these approaches generally fuse
both feature and spatial information to produce more compact and
well-dened regions [
10
]. The graph-based approaches dene a
graph whose vertex corresponds to a region, and weight of edge is
dened as the likelihood to segmentation. A graph is separated into
components according to the cost function of vertices and edges.
Graph-based methods require high computation complexity and
consequently disables their use in real-time applications. Image fea-
ture extraction in combination with Machine Learning techniques
is widely used is several applications. In [
8
], Low-Level features,
extracted from hyper-spectral images, are used as partial inputs
to semi-supervised ML algorithms for surveillance purposes. As
mentioned, high detection ability of the algorithms was conrmed.
Unsupervised learning in healthcare studies commonly utilizes
segmentation methods, which allows the algorithms to nd hidden
patterns or groupings in data [
12
]. Extracting new features from
the original features is useful for reducing the dimension of feature
space and achieving better performance. According to [
13
], it is
stated that unsupervised ML techniques can be used for the iden-
tication of certain ROIs of an image. The corresponding results
show that specic, rough regions with similar characteristics can
be grouped in the same cluster, increasing the eciency of the
algorithm. In literature, various segmentation algorithms are uti-
lized to identify typical regional peak plantar pressure distributions
of a diabetic foot [
2
] using thermal images. However, extracting
features for diabetic foot monitoring using RGB images is not a
well-studied domain.
3 PROPOSED METHODOLOGY
Image segmentation is the technique of dividing an image into
homogeneous, disjoint, and meaningful parts and extracting the
parts of interest. The main idea of this approach is the initial feature
extraction from the original image to detect ROIs and then, the
implementation of clustering techniques to segment these regions.
In this section, we describe the applied pre-processing tech-
niques, for the construction of the Low/High-Level features and
analyze, briey, the applied clustering approaches. The whole pro-
cedure is depicted in Figure 1. Initially, the original image is passed
through the feature extraction layer. The extracted LL/HL features
are packed into a multi-channel array, containing the original image
as well. This image is used as an input to the selected clustering
method, which produces the corresponding labels of the output.
These labels are used for the reconstruction of several 2-dimensional
images, containing the grouping information. All the distinct clus-
ters are considered to be the potential regions of interest and thus
they are all combined into a single image at the nal step.
3.1 Low level features
Image pre-processing is performed on the images to improve the
eectiveness of the proposed algorithm and accurate detection of
the ROI. Image enhancement and ltering techniques have been
applied, to eliminate the eects of noise and illumination in the
scanning area. The main lters that were used for this initial ap-
proach are listed below:
(1)
Entropy lter: For an image, local entropy is related to the
complexity contained in each neighborhood, typically de-
ned by a structuring element. Entropy lter calculates the
entropy of the neighborhood and assigns that value to the
output pixel. Entropy function gives a value that represents
level of complexity in a certain section of an image.
(2)
Canny edge detector: The Canny edge detector is an edge
detection operator that uses a multi-stage algorithm to de-
tect a wide range of edges in images. Canny edge detection
technique is used to extract useful structural information
from dierent vision objects and dramatically reduce the
amount of data to be processed. It has been widely applied
in various computer vision systems. Canny edge detection
2
609

Unsupervised diabetic foot monitoring techniques PETRA ’22, June 29-July 1, 2022, Corfu, Greece
algorithm is one of the most strictly dened methods that
provides good and reliable detection. Owing to its optimality
to meet with the three criteria for edge detection and the
simplicity of process for implementation, it became one of
the most popular algorithms for edge detection.
(3)
Low-pass lter: A low pass lter is a smoothing method that
decreases the disparity between pixel values by averaging
the nearby pixels.
In Figure 2, the application of the lters described above, are de-
picted in the original image. In this case, all the lters seem to
dierentiate important aspects of the ROI. The left most image is
the original one. The second image depicts the result of the low
pass lter applied on the original one. The third image presents the
result of entropy lter. The fourth image shows the Canny edge
detector outcome.
(a) (b) (c) (d)
Figure 2: Indicative outcomes when using the proposed low
level features over an RGB image (a), low pass lter (b), en-
tropy lter (c) and Canny edges (d).
3.2 High-Level features
In this scenario we investigate a High-Level, feature-based, pixel-
level, clustering segmentation approach. The idea lies in creating
coherent regions of neighboring pixels in a way that any com-
plication is attributed to a common cluster. This can happen by
exploiting the features generated by a pre-trained Deep Learning
(DL) model.
The HL feature extraction is based on the utilization of an exist-
ing DL model, trained on similar data with our case. In this scenario,
we will utilize the extracted features from a convolutional denoising
stacked auto-encoder (SAE), trained over the same dataset, using
as input Gaussian induced noise images. The feature values from
any intermediate level of this model can serve the cause for the
demonstration purposes.
Figure 2.a demonstrates an input image. This image was resized
and fed as input to a pre-trained convolutional denoising SAE.
The image is transformed using various operators as it progresses
through the hierarchical structure. Depending on the number of
kernels in the layer, the original image is transformed to many
smaller ones, with each of them emphasizing dierent areas and
shapes.
Figure 3 demonstrates the high level extracted features, as gray
scale images. We have 64 images as the number of kernels in the
corresponding layer, where the extraction occurred. Brighter color
indicates ROIs and darker scales background context. These fea-
tures were the outcome of the second layer in the convolutional
denoising SAE model. Some images are similar to traditional ap-
proaches like edge detection. Other images emphasize on specic
skin irregularities/discoloration.
Figure 3: High level features extracted.
4 EXPERIMENTAL SETUP
4.1 Dataset description
In this paper, we have utilized the DFUC 2020 dataset [
4
]. This
dataset is publicly available for non-commercial research purposes
only and can be obtained by emailing a formal request to the authors.
The DFUC 2020 dataset consists of 4,000 images, with 2,000 used
for the training set and 2,000 used for the testing set.
4.2 Performance metrics
Two dierent types of performance scores were considered: a) clus-
tering based, as described in 4.2.1 and b) object detection based,
described in the next paragraph. The former case evaluates the
cohesion of the clusters and the latter the applicability for ROI
detection, in an unsupervised way.
4.2.1 Clustering based metrics. Formally, a cluster analysis can
be described as the partitioning of a number of classication ob-
jects in
K
groups or clusters
{Ck},k=
1
, ..., K
. Given
N
objects
X={x1, ..., xN}
, where
xi j
denotes the
j−th
element of
xi
. The
grouping of all objects
xi,i=
1
, ..., N
in
K
clusters can be dened
as follows:
wki =(1i f f x i∈Ck
0otherwise (1)
The Calinski Harabasz Index (CHI)[
3
] is dened according to
the following equation:
CH I (k)=
TB
K−1
Tw
N−K
(2)
where TBis dened as:
TB=
K
Õ
k=1
|¯
CK|∥CK−¯
x∥(3)
3
610

PETRA ’22, June 29-July 1, 2022, Corfu, Greece Tzortzis, et al.
Splitter
MAX
The predicted masks The ground truth mask
Intersection over union calculation
Clustered Img with Bounding Boxes
The best predicted mask
Figure 4: The evaluation process for the ROI identication performance.
and Twis dened as:
TW=
K
Õ
k=1
N
Õ
i=1
wki ∥xi−¯
CK∥2(4)
Tw
starts at a comparably large value. With increasing number
of clusters
k
, approaching the optimal clustering solution in
K
groups, the value should signicantly decrease due to an increasing
compactness of each cluster. As soon as the optimal solution is
exceeded an increase in compactness and thereby a decrease in
value might still occur; this decrease, however, should be notably
smaller. Calculated for each possible cluster solution, the maximum
CHI value indicates the best cluster partitioning of the data.
The Davies-Bouldin Index (DBI) [
5
] is an internal evaluation
scheme, where the validation of how well the clustering has been
done is made using quantities and features inherent to the dataset.
DBI is dened as follows:
DBI (k)=1
K
K
Õ
k=1
RK(5)
where Rkis dened as:
RK=max Sk+Sj
dk j !,j=1,2, . . ., K,j,K(6)
dk j
is a distance function, dened as
dk j =∥¯
xk−¯
xj∥
and
Sk
is
dened as:
SK=1
ÍN
i=1wki
N
Õ
i=1
wki ∥xi−¯
xk∥(7)
All the above equations assume that
k∈ [
1
,K]
. For each cluster
Ck
an utmost similar cluster—regarding their intra-cluster error
sum of squares—is searched, leading to
Rk
. The index then denes
the average over these values. In this case, the minimum index
value corresponds to the best cluster solution.
The silhouette value is a measure of how similar an object is to
its own cluster (cohesion) compared to other clusters (separation).
The silhouette ranges from -1 to 1, where a high value indicates that
the object is well matched to its own cluster and poorly matched
to neighboring clusters. If most objects have a high value, then
the clustering conguration is appropriate. If many points have a
low or negative value, then the clustering conguration may have
too many or too few clusters. For each datum
xi
, let
α(xi)
be the
average dissimilarity (distance) of
xi
with all other data within the
same cluster
Ck
. Let
b(xi)
be the lowest average dissimilarity of
xi
to any other cluster
Cl
,
l,k
, of which
xi
is not a member. We now
dene a silhouette as:
s(xi)=b(xi) − a(xi)
max(b(xi),a(xi)) (8)
Thus,
s(xi) ∈ [−
1
,
1
]
; values close to one indicate that the datum
xi
is appropriately clustered at
Ck
. The average silhouette value
over all data, i.e.,
¯
s=1
nÍn
i=1s(xi)
, is another measurement for the
quality of the generated clusters.
4.2.2 Object detection based metrics. Figure 4 demonstrates how
we calculate the Intersection over Union (IoU) scores for a clustering
problem. The clustered image, containing all the predicted bounding
boxes, is given as input. The splitter mechanism divides the input
image object into distinct masks according to the detected bounding
boxes. These masks are compared to the ground truth masks of the
dataset using the IoU score. The predicted mask with the highest
IoU score is forwarded as output of the evaluation process pipeline.
4.3 Experimental results
In this section, the results of the proposed method are presented,
in terms of grouping problematic skin areas in the same cluster.
We demonstrate how well the proposed scheme performed, over
the test set. Recall that the clustering was based on LL features.
Except from the average performance scores, we considered the
minimum and maximum values, which can indicate uctuations in
performance.
The top-left graph of Figure 5 demonstrates the CHI scores for
various clustering approached, using LL features. The higher the
CHI score, the better the corresponding approach. In this scenario,
K-means appears to be the best clustering technique. Even the worst
observed CHI value, when using K-means, surpasses the alterna-
tives. The worst technique seems to be the Meanshift algorithm.
Similar results are shown in the top-left graph of Figure 6, which
depicts the CHI scores for the several clustering algorithms, using
4
611

Unsupervised diabetic foot monitoring techniques PETRA ’22, June 29-July 1, 2022, Corfu, Greece
HL features. In this case, there is an improvement regarding the
Spectral clustering algorithm.
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
Kmeans Meanshift Spectral
Performance score
Clustering technique
CHI score Max CHI score Min CHI score
0
0,2
0,4
0,6
0,8
1
1,2
1,4
1,6
1,8
2
Kmeans Meanshift Spectral
Performance score
Clustering technique
DBI score Max DBI score Min DBI score
0
0,2
0,4
0,6
0,8
1
1,2
1,4
1,6
1,8
2
Kmeans Meanshift Spectral
Performance score
Clustering technique
DBI score Max DBI score Min DBI score
Figure 5: a) CHI scores for various clustering approaches,
using LL features. b) DBI scores for various clustering ap-
proaches, using LL features. c) Silhouette scores for various
clustering approaches, using LL features
The top-right graph of Figure 5 illustrates how the DBI scores
varied, when using dierent clustering approaches, when applying
LL features. The lower the DBI score, the better the corresponding
technique behaves. According to this metric, Meanshift algorithm
appears to be the best case. At the same time, this technique demon-
strates the narrowest range between minimum and maximum ob-
served values. Spectral clustering scored the worst results. On the
other hand, this algorithm seems to be more appropriate when
using HL features, as it is shown in top-right graph of Figure 6. It
Figure 6: a) CHI scores for various clustering approaches,
using LL features. b) DBI scores for various clustering ap-
proaches, using LL features. c) Silhouette scores for various
clustering approaches, using LL features
corresponds to the lowest average DBI value and it appears to have
the smallest deviation around these values.
The bottom graph of Figure 5 presents the Silhouette scores for
various clustering approaches, using LL features. The closer the
Silhouette score is to 1, the better the corresponding approach is.
In this case, Meanshift appears to be the best approach, with an
average score of 0.375. Similar results occur when HL features are
used. In this case, K-means appears to achieve the best Silhouette
score. Please note than the highest observed score for both cases
was 0.5. This is an indication that all of the proposed schemes
provide limited information to the physician/patient.
Figure 7 provides some indicative clustering results for image
ID 100047. In this scenario LL features were used as inputs for the
proposed clustering algorithms. The rst image of the rst row
shows the best predicted bounding box using k-means clustering
algorithm. The second and third image in the same row demonstrate
the results for Meanshift and Spectral clustering, respectively. The
images in the bottom demonstrate where these bounding boxes
should be placed. The fourth image is the ground truth. Meanshift
appear to be more precise in cluster creation but spectral clustering
mapped better the ulcer region.
5
612

PETRA ’22, June 29-July 1, 2022, Corfu, Greece Tzortzis, et al.
(a) (b) (c) (d)
Figure 7: Clustering results and the corresponding ROI for
image ID 100047, using LL features. The columns of the grid
display labeled with (a), (b), (c), (d) correspond to K-means,
Meanshift, Spectral and ground truth respectively.
Figure 7 provides some indicative clustering results for image
ID 100059. In this scenario LL features, described in subsection 3.1,
were used as inputs for the proposed clustering algorithms. Cluster-
ing outcomes are fuzzy and cannot be interpreted. The localization
appears to be better, but this is attributed to the adopted calculation
mechanism, as described in subsection 4.2.1.
(a) (b) (c) (d)
Figure 8: Clustering results and the corresponding ROI for
image ID 100059, using HL features. The columns of the grid
display labeled with (a), (b), (c), (d) correspond to K-means,
Meanshift, Spectral and ground truth respectively.
Regarding the average value and the divergence of the execution
time and the IoU score, the corresponding results are presented
in Figure 9. The left graph of Figure 9 depicts the execution time,
in seconds, for the proposed segmentation schemes. In this case,
K-means was the fastest method for both implementations (HL
and LL). Mean shift was a lot slower in the HL implementation
and almost approached the time of K-means in the LL implemen-
tation. Spectral clustering technique seems to be slower in both
implementations from K-Means, yet, faster from mean shift in HL
approach.
The right graph of Figure 9 depicts the IoU scores for the pro-
posed segmentation schemes. The higher the IoU score, the better
the accuracy of the technique. In this case, K-means and Meanshift
seem to be the best two methods for the HL implementation (al-
most the same score). K-means, however, is slightly better because
it gives higher minimum score. Spectral method is the best method
for the LL implementation, with the highest score value, while
K-means is slightly better than the Meanshift.
0
20
40
60
80
100
120
DL_features LL_features DL_features LL_features DL_features LL_features
Kmeans Meanshift Spectral
Execution time (seconds)
Combinatory approach
Avg value Max value Min value
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
DL_features LL_features DL_features LL_features DL_features LL_features
Kmeans Meanshift Spectral
Performance score
Combinatory approach
IoU Max IoU Min IoU
(a) (b)
Figure 9: (a) Execution times (seconds) for the proposed seg-
mentation schemes - (b) IoU scores for the proposed segmen-
tation schemes.
5 CONCLUSIONS
In this paper, we provide a set of Machine Learning tools for image
segmentation in medical images of diabetic foot, to detect ROIs
and to extract meaningful features to boost model performance in
DFU detection and monitoring. The imagery data that are ingested
into the unsupervised Machine Learning algorithms, have been
previously processed using either Low-Level descriptors (ltering
methods) or High-Level features using Deep Learning approaches.
There are two major outcomes from this study: a) unsupervised
techniques inappropriateness and b) High-Level feature capabilities.
In the former case, results suggest that unsupervised techniques
are time consuming and cannot provide adequate results that could
help physicians/patients during monitoring/treatment phase. In
the latter scenario, we see that pixels corresponding to problematic
areas can be grouped together, with relatively good performance.
This is a strong indication that detection and localization of the
ROI is feasible using supervised learning techniques.
Thus, in the future, we will utilize the knowledge gained from
this work, in an eort to propose novel and sophisticated (semi)-
supervised approaches.
ACKNOWLEDGMENTS
The work in this paper has been supported by the H2020 Phooton-
ics project: “A Cost-Eective Photonics-based Device for Early
6
613

Unsupervised diabetic foot monitoring techniques PETRA ’22, June 29-July 1, 2022, Corfu, Greece
Prediction, Monitoring and Management of Diabetic Foot Ulcers”
funded under the ICT H2020 framework and the grand agreement
no. 871908.
REFERENCES
[1]
N Arteaga-Marrero, A Hernández, E Villa, S González-Pérez, C Luque, and J.
Ruiz-Alzola. 2021. Segmentation Approaches for Diabetic Foot Disorders. Sensors
75 (2021). https://doi.org/10.3390/s21030934.
[2]
Craig J Bennetts, Tammy M Owings, Ahmet Erdemir, Georgeanne Botek, and
Peter R Cavanagh. 2013. Clustering and classication of regional peak plantar
pressures of diabetic feet. Journal of biomechanics 46, 1 (2013), 19–25.
[3]
Tadeusz Caliński and Jerzy Harabasz. 1974. A dendrite method for cluster analysis.
Communications in Statistics-theory and Methods 3, 1 (1974), 1–27.
[4]
Bill Cassidy, Neil D Reeves, Joseph M Pappachan, David Gillespie, Claire O’Shea,
Satyan Rajbhandari, Arun G Maiya, Eibe Frank, Andrew JM Boulton, David G
Armstrong, et al
.
2021. The DFUC 2020 dataset: Analysis towards diabetic foot
ulcer detection. touchREVIEWS in Endocrinology 17, 1 (2021), 5.
[5]
David L Davies and Donald W Bouldin. 1979. A cluster separation measure. IEEE
transactions on pattern analysis and machine intelligence 2 (1979), 224–227.
[6]
Anastasios Doulamis, Nikolaos Doulamis, Aikaterini Angeli, Andreas Lazaris, Siri
Luthman, Murali Jayapala, Günther Silbernagel, Adriane Napp, Ioannis Lazarou,
Alexandros Karalis, et al
.
2021. A Non-Invasive Photonics-Based Device for Mon-
itoring of Diabetic Foot Ulcers: Architectural/Sensorial Components & Technical
Specications. Inventions 6, 2 (2021), 27.
[7]
Kostas Haris, Seram N Efstratiadis, Nikolaos Maglaveras, and Aggelos K Kat-
saggelos. 1998. Hybrid image segmentation using watersheds and fast region
merging. IEEE Transactions on image processing 7, 12 (1998), 1684–1699.
[8]
Konstantinos Makantasis, Eftychios Protopapadakis, Anastasios Doulamis, and
Nikolaos Matsatsinis. 2015. Semi-supervised vision-based maritime surveillance
system using fused visual attention maps. Multimedia Tools and Applications 75
(03 2015). https://doi.org/10.1007/s11042-015-2512- x
[9]
Sokratis Makrogiannis, George Economou, and Spiros Fotopoulos. 2005. A region
dissimilarity relation that combines feature-space and spatial information for
color image segmentation. IEEE Transactions on Systems, Man, and Cybernetics,
Part B (Cybernetics) 35, 1 (2005), 44–53.
[10]
Jianbo Shi and Jitendra Malik. 2000. Normalized cuts and image segmentation.
IEEE Transactions on pattern analysis and machine intelligence 22, 8 (2000), 888–
905.
[11]
Wenbing Tao, Hai Jin, and Yimin Zhang. 2007. Color image segmentation based
on mean shift and normalized cuts. IEEE Transactions on Systems, Man, and
Cybernetics, Part B (Cybernetics) 37, 5 (2007), 1382–1389.
[12]
Jack Tulloch, Reza Zamani, and Mohammad Akrami. 2020. Machine learning in
the prevention, diagnosis and management of diabetic foot ulcers: a systematic
review. IEEE Access 8 (2020), 198977–199000.
[13]
Charalampos Zafeiropoulos, Ioannis Tzortzis, Ioannis Rallis, Eftychios Protopa-
padakis, Nikolaos Doulamis, and Anastasios Doulamis. 2021. Evaluating the
Usefulness of Unsupervised monitoring in Cultural Heritage Monuments. (07
2021).
[14]
Hui Zhang, Jason E Fritts, and Sally A Goldman. 2005. A fast texture feature
extraction method for region-based image segmentation. In Image and Video
Communications and Processing 2005, Vol. 5685. International Society for Optics
and Photonics, 957–968.
[15]
Yan Zhang and Xiaoping Cheng. 2010. Medical image segmentation based on
watershed and graph theory. In 2010 3rd International Congress on Image and
Signal Processing, Vol. 3. IEEE, 1419–1422.
7
614