Signal Processing - July 2016 - 44

only for capturing, storing, and transmitting skin lesion images to a remote server [26]. As discussed for the wound assessment apps, this requires a stable data link and creates privacy
and security challenges that are usually ignored by the developers, who assume that they will be handled by the (cloud)
computing provider.
Nonetheless, a few systems perform the analysis of smartphone-captured or dermatoscopic images directly on a mobile
device. For example, Wadhawan et al. [27] proposed an ondevice library for melanoma detection based on a bag-of-features framework. They showed that the most computationally
intensive and time-consuming algorithms (image segmentation and image classification) can achieve an accuracy and execution speed comparable to a desktop computer.
On-device processing of images captured by a smartphone
has been reported in [28] and [23]. The system proposed in [28]
uses a basic thresholding method for segmentation, followed
by classification based on the standard color feature and border
features, but the results have low accuracy (66.7%). In contrast,
Doukas et al. [23] focused on the system integration, without
describing in detail the features used to reach the reported
accuracy of 77.06%.
In [7], a fully on-device melanoma diagnosis system is proposed that employs a combination of skin detection and edge
localization for fast lesion segmentation. The segmented lesion
is then used for feature extraction. The features used are motivated by the original features defined by dermatologists, also
known as the ABCD feature set: asymmetry, border irregularity, color and texture variation, and diameter [7]. To cope
with the limited resources available on the mobile device, the
authors devised a new feature selection mechanism that takes
into account the coordinate of the feature values to identify
more discriminative features and efficiently process them.
Given the feature set F and the class label L , the feature
selection problem is to find a set S 1 F ^| S | 1 | F | h such
that it maximizes the relevance between L and S, usually
characterized in terms of MI. The hybrid criterion used for
feature selection combines NMIFS and average neighborhood
margin (ANM) maximization schemes in a single unified criterion defined by the following relation [7]:
U fi = a · Q fi + (1 - a) · ;MI (L, fi) - 1
|S |

fs ! S

NMI (fi, fs)E,

(6)

where $ is the cardinality of the set, Q f is the quality of
feature f (defined below), MI is the mutual information that
measures the relevance between two discrete random variables
X and Y (with alphabets X and Y ), NMI is the normalized MI,
and a ! [0, 1] is a weight factor that controls the influence of
the ANM and MI in the proposed hybrid criterion.
The quality of the feature f (which is derived from ANM)
is defined as
n

Qf =

i= 1

n ei

t= 1

fi - ft
n ei

n oi

j= 1

fi - f j
n oi

o ,

where, for each sample i, n oi is the set of the most similar samples that are in the same class with sample i, and n oi is the set
of the most similar samples that are not in the same class with
i. The set of selected features is extracted and used by a classifier array to produce the detection results, reporting over 80%
sensitivity and specificity.
A comparison of the existing smartphone apps (available
for Android and iOS platforms) and proposed approaches is
presented in Table 5 in terms of use of automated decisionmaking (i.e., machine learning), image modality, accuracy, test
bed size, and on-device/on-server processing. As indicated by
the reported accuracies, test bed size, and image modalities,
smartphone-based skin cancer diagnosis apps are not developed enough to be used even by experts and, therefore, are not
ready to be used for assisted-living scenarios.

Discussion
Smartphones are used in a wide range of situations in everyday
life and thus provide a unique opportunity to help users daily.
Offering several different sensors and computational power,
smartphones have proved to be a promising tool for assisted
living in both cities and isolated rural areas, with apps ranging
from human-computer interfaces and augmented reality apps
for disabled users to health-monitoring and fitness-tracking
devices. Smartphone-based health-monitoring apps compose
one of the main tracks of assisted-living apps, which have
drawn increased attention due to the growth of the aging population as well as eHealth ecosystems. However, the challenge
of developing a well-designed app for health monitoring should
not be underestimated. To achieve the desirable accuracy, performance, and usability, techniques from different fields, such
as interface design, sensor fusion, signal processing, image processing, and machine learning, should be properly combined to

Table 5. A comparison of representative existing systems used for skin cancer diagnosis.

Abuzaghleh et al. [26]
Wadhawan et al. [27]
Ramlakhan et al. [28]
Doukas et al. [23]
Do et al. [7]
Lubax

Machine
Learning
No
Yes
Yes
Yes
Yes
Yes

Accuracy
90%
79.26%
66.7%
77.06%
84%
Not Available

(7)

Processing
Server
Local
Local
Local/Server
Local
Local

IEEE SIgnal ProcESSIng MagazInE

Image Modality
Dermoscopy
Dermoscopy
Visible Light
Dermoscopy
Visible Light
Visible Light

July 2016

Data Set
200
347
83
3,000
184
>12,000

Operating System
Platform
iOS
iOS
iOS
Android
Android
iOS/Android

Table of Contents for the Digital Edition of Signal Processing - July 2016