Skin Cancer Detection Application

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 03 | Mar 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 350
Skin Cancer Detection Application
Tushar Gharge [1], Narayan Gupta [2], Krutik Raut [3], Prof. Varsha Salunkhe [4]
[1], [2], [3]Student, Department of Computer Engineering, Atharva College of Engineering, Mumbai
[4]Professor, Department of Computer Engineering, Atharva College of Engineering, Mumbai
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - The Project Entitled As “Skin Cancer Detection
Application”. Skin cancer is regarded as one of the most
dangerous types of cancer seen in humans. Every year, tens
of thousands of people in the United States are diagnosed
with cancer. As the disease spreads, the rate of survival
drops dramatically. Melanoma, Basal and Squamous cell
Carcinoma are the most common types of skin cancer, with
Melanoma being the most unpredictable. Early detection
and identification of the type of cancer can aid in its cure.
Many technologies have demonstrated that AI and
computer science can play a key role in picture scanning.
We describe a mobile technique for detecting Melanoma
Skin Cancer utilising Image Processing and basic portable
equipment in this technical study. The cancer image is fed
into the system, which analyses it using unique image
processing algorithms to determine the existence of skin
cancer. The cancer picture analysis software looks for
asymmetry, border, colour, diameter, and other Melanoma
properties, as well as size, frequency, and form, to identify
the image and feature phases. The scanned feature
parameters are used to distinguish between normal skin
and melanoma cancer lesions in the image.
Key Words: Melanoma, TensorFlow, Dermatology, Image
Processing, Transfer Learning,
1. INTRODUCTION
In the age of apparatuses, machine learning is finding its
application in all fields and clinical diagnosis is the newest
calculation to it. . Emerging machine learning (ML)-based
analytical tools are exciting to build, but they must be
chosen carefully. To improve diagnostic accuracy, choose
the correct decision-making system. Face identification,
fingerprint recognition, tumour detection, and
segmentation are just a few of the applications of image
processing and machine learning. Fungal infection,
bacteria, allergies, and viruses, among other things, can
cause skin problems.
The project's purpose is to create and test robust,
automated textural feature extraction and selection
algorithms, as well as learning algorithms for identifying
pigmented skin lesions, in order to improve skin cancer
diagnosis in primary care settings. The goal of this
initiative is to enhance the detection and treatment of skin
cancer. One of the project's main objectives is to develop
an interactive support tool with a simple user interface to
assist primary care practitioners in accurately screening
pigmented skin lesions in order to improve early
detection, reduce unnecessary hospital referrals, and
reduce the number of malignant melanomas missed.
The impact of the proposed project is twofold significantly
improved understanding of correspondence between
automated extracted features The ABCD rule of thumb and
computerised automatic identification (such as size, shape,
colour, and texture information) physicians used to pre-
screening the early skin cancer; and advances integration
of feature extraction and classification in bio-medical
imaging processing. The overarching goal is to provide
transparent support for skin cancer screening by primary
care physicians based on advanced bio-medical image
processing and decision-making techniques.
1.1 MOTIVATION OF PROJECT
The primary motivation that drives the project is to use
the advanced image classification technology for the well-
being of the people. Computer vision has made good
progress in machine learning and deep learning that are
scalable across domains. The model's improved accuracy
and efficiency may help detect melanoma in its early
stages and prevent unneeded biopsies.
1.2 BASIC CONCEPT
We would gather images from user's devices via camera or
system storage and identify skin cancer diseases like
melanoma to their closest prediction in this application. It
would give a description and symptoms of the condition
after successful identification. The user could also look up
the contact information for nearby doctors.
It would also provide chart data as well as approximations
of other nearby conditions via a user-friendly interface,
allowing consumers to be directed to specific skin disease
results.
2. BACKGROUND STUDY
Sheha, Mariam A. On a collection of dermoscopy photos,
Mai S.Mabrouk and Amr Sharawy [1] proposed an
automated method for melanoma diagnosis. So, while the
first exercise, Automatic iteration counter, is faster, the
second practise, Default iteration counter, is more precise,
with a precision of 100% for the training set and 92
percent for the test set.
Jonathan Blackledge et al. [2] proposed the Moletest online
skin growth screening framework.
The half-examination of a photograph in terms of its
fractal structure and the fractal qualities that characterise
it.

So, which of these two perspectives, or a combination of
both, has been used to describe a handling and image test
engine that is unique in its customary technique and
entirely nonexclusive in terms of the applications to which
it may be connected.
In light of form descriptors, Messadi M. et al. [3] developed
an interpretable order technique for skin cancers in
dermoscopic images. Their research demonstrates the use
of a fuzzy rule-based classifier to distinguish between
melanoma and other cancers. A flexible Neuro Fuzzy
inference System (ANFIS) has a specific purpose in mind:
to uncover the fuzzy rules that prompt the correct
classification. The technique used in previous stages of
development was to reduce the impact of small structures,
hairs, bubbles, and light refraction. In the second step, an
unsupervised lesion segmentation algorithm is proposed.
In a transparent way, repeating thresholding is linked to
the instate level specified. They've also addressed the
requirement to focus all the ascribes employed to add to a
depiction strategy that enables professionals to make the
best decision.
The "ELM 7-point agenda," developed by G. Di Leo et al.
[4], is a new analytic approach that describes an
arrangement of seven components, including texture
parameters, that convey the malignancy of a lesion. It has
been shown to be faster and more precise than the
traditional ABCD criteria in the detection of melanoma.
Capdehourat et al. [5] from Germany suggested a machine
learning method for dealing with order melanocytic
lesions in hazardous and generous dermatoscopic images.
There are 433 benign sores and 80 benign lesions in the
image database. Following a step of image pre-processing
that includes hair evacuation filtering, each image is then
fractured using well-known image segmentation
techniques.Yogendra Kumar Jain et al.
[6] focuses on the development of a skin cancer screening
framework that may be used as part of routine practise by
non-specialists to distinguish between normal and
abnormal instances. Highlight Discovery and Order
Procedure are two steps in the advancement process. The
components are separated using wavelet change to
decompose images into multiple recurring sub-groups. As
a result, the characterization framework is based on the
use of a Probabilistic Neural System and a Grouping
Classifier.
Omar Abuzaghleh et al. [7] suggested an ingenious and
entirely practical advanced mobile phone-based
programme for determining the type of early disease and
its prevention. The first segment is a constant alarm to
provide customers with help in preventing skin smoulder
caused by sunshine, and the second segment is a new
counted statement to know the most of the possibility for
skin to points is presented along these lines. The proposed
framework is effective, according to the output data,
grouping typical, atypical, and melanoma images with 96.3
percent, 95.7 percent, and 97.5 percent accuracy,
respectively.
V. Jeya Ramya et al. [8] proposed an automated approach
with typical and anomalous classes for skin cancer
identification. The great division execution is
accomplished by active contour segmentation. In an order
approach with two classifications (threatening and
favorable sores), an affectability of 90%, precision of 95%
and a specificity of 85% is watched.
3. IMPLEMENTATION
Fig-1: System Architecture
3.1 DATA COLLECTION
We start by gathering the best dataset containing
photographs of skin conditions. We started by looking for
melanoma-related photos on Kaggle and dataverse. While
searching for skin disorders, there were numerous
datasets with a large number of good photographs and
models, so we choose the one with the most likes or
recommends. We discovered Kaggle's SIIM-ISIC
Melanoma Classification, which included over ten GB of
data, and Harvard's data verse HAM10000 images part 1
& HAM10000 images part 2 dataset image when
searching for the top datasets. We chose these two
datasets because of the enormous number of datasets and
recommendations. Normal and day-to-day living duties,
educators, and psychiatrists might benefit from knowing
the distinction between benign and malignant types.
Choosing the best photographs from the two datasets,
merging them, and categorizing them. Data was framed
and stored in csv files from the obtained photos and
datasets.
3.2 DATA CLEANING AND PREPROCESSING
Maintaining the data is very important so that bad data
doesnot coincide with the training and modelling of our
main and working model. We focus to pull out the
important data that is required for the process. Before

beginning with the cleaning, we concatenated the two sets
and added a special column to indicate the image is from
which type of cancer it is. The firststep in cleaning was to
reduce the number of unnecessary columns from out data
framed sets. Most of the sets had aboutnumerous amounts
of columns, of which only few we considered were
important. This included the title of the disease, date of
the content, resolution of the content, and the probability
or confidence column that we added.
Each image must be unique and clear in way of visibility,
while maintaining it’s original dimensions and data. For
even better results, .dcm files: DICOM files were also
added. Most of the images were saved in the "Digitals
Image and Communication in Medicine" format. It
contains an image from a medical scan, such as an
ultrasound or MRI + information about the patient. In
order to prepare the sets for our classifier pre-processing
of these images was required. This includes proper
lighting, removing any other visual discrepancies. After
the data was pre-processed and cleaned, it was added into
our column in of our csv files.
3.3 DATA MODELLING
To decide on training model, we used TensorFlow lite
transfer learning with MobileNet-v2 architecture model.
training a TensorFlow Lite model with your own custom
dataset decreases the number of trained data required and
will shorten the overall time. So, the pre-trained neural
network would help you to create your own classes, and so
you can sort of picture that your class are becoming the last
layer or step of our neural net.
MobileNet-v2 is a CNN that is 53 layers deep. People
can load a pretrained version of the network trained on
more than a million images from the ImageNet database.
The pre-trained CNN can classify images into most of the
real time objects categories, such as keyboard, mouse,
pencil, and many test data. Such kind of result is the which
has the network and learned rich feature representations
for a wide range of images. The output has an image input
Fig-2: Training model
size of 224-by-224.
Fig-3: MobileNet-v2
In the context of our project for training model, we would
use the teachable machine’s image classification to train
the model for ease of having better accuracy and result. We
would do the following steps:
Added the cleaned and preprocessed dataset for
classification as group of classes based on the disease name
of the respective data. After all the data was uploaded, it
was ready for the next process of training.
For training the model with nearest accuracy possible, we
tried to adjust the models’ parameters of epochs (for
number of times dataset fed for training), batch size (set of
samples fed in each iteration), learning rate (rate to adjust
the models understanding) accordingly.
So, after going multiple trial and run cases, we achieved
our accuracy highest accuracy at epochs of 89, batch size
of 128 and learning rate of 0.0001. After working our
model on multiple permutations of the above parameters,
we saved the result match and tried to deploy it for the
testing stage.
The optimized model scored well on our test set, scoring
an AUC score of 0.65 ~ 0.75. We would try proceed to
know our model a bit better before making final
estimation and recommendations.

4. LOADING AND TESTING
The final phase was to load and test whether our trained
model submitted on real time devices works accurately or
not. After training, the trained model was converted into
tflite quantized model as the extension of. tflite. After the
model was converted, it was loaded and integrated into
Flutter code and was made to run on the live images from
the devices. The model was made to run on threshold of
0.5, image mean and standard of 127.5.
Testing was done on many real time skin cancer diseases
of melanoma from many numbers of individuals and
patients. We achieved the accuracy of 0.60 ~ 0.80 approx.
The images taken with better illumination and high
resolutions showed the higher accuracy of around 0.70
and above, while the images having lower pixels or
resolutions and less illumination had little lower result
range in between of 0.40 and 0.60. Overall, the model
worked accurately from the portable device like mobile
trained from the MobileNet-v2 architecture. Response
time was very quick and was able to identify the diseases
even in the low-end devices.
Fig-4: Loading model
Our gives average of 70.6 percent accuracy when looking
at regular entries (i.e. is not including the variance of other
diseases outside of melanoma). Final the level of accuracy
means that the model might have the potential to be
generalize for usage in day to day like schools, rural areas
and even for just verifying the disease for anyone who
download the application.
Fig-5: Testing on real time images
5. RESULTS
At last, the modelling and testing phase we can conclude
that our classifier has an above average rating in terms of
being correct and providing accurate results. This means in
the future such a system can be implemented to check if a
person or patient has been experiencing any kind of
melanoma cancers that might lead to possible even bigger
problems. Such a system can also be implemented on other
web applications, IOS & android devices as well as
embedded devices. In the end our model was able to
achieve close to 70 percent rate of success, with more
research and studies this rate can be increased even
further and provide more accurate and dependable results.
The main issue of the work was that a lot of pre-processing
and collecting the data, which was required even on the
sets on which the models were tested on to provide near
results. While training more and more of our model from,
datasets we would be able to know how our output would
be able to change in different scenarios.

6. DESIGN DETAIL
6.1 E-R DIAGRAM
Fig-6.1: E-R diagram
6.2 USE CASE DIAGRAM
Fig-6.2: Use case diagram
7. FUTURE SCOPE
This idea and approach that we have implemented does
not end here. There can be a future scope for development
we can extend this project into multiple platforms.
Particularly, we can incorporate additional algorithm
together to make is more appealing.
One other scope of improvement in expanding the idea
hiding the data into a video format so that the application
can be more secure so that different sector can use this to
secure the way of communication between the user and
themself. As specially, since the sensitive data can be
stored without anyone knowing about it. Such that storing
sensitive data won’t be a problem it as we add higher level
of security.
8. CONCLUSION
Skin cancer detection is very hard to for most of the
people to detect without visiting doctors. Main project of
our diagnosis requires experience of data, in which early
stages may look identical to benign. Detecting the disease
with such tool for less experience physicians. Our model's
project is an automated melanoma classification system
that was applied to dermoscopy pictures to aid in the early
diagnosis of malignant melanoma and melanocytic nevi
lesions. With main advantage that it contrasts with other
methods in medical image analysis segmentation process
is avoided using mobile analysis. In this project we had
created a mobile application to classify skin cancer. This
method consists of data classification, image processing
and image classification using Mobile net model. Using this
technique, we can classify images at nearest probability.
9. ACKNOWLEDGEMENT
We owe sincere thanks to our college Atharva College of
Engineering for giving us a platform to prepare a project
on the topic "Skin cancer Detection application" and would
like to thank our principal Dr. Shrikant Kallurkar for
instigating within us the need for this research and giving
us the opportunities and time to conduct and present
research on the topic. We are sincerely grateful for having
Prof. Varsha Salunkhe as our guide and Dr. Suvarna
Pansambal, Head of Computer Engineering Department,
and our project coordinators during our research. So, at
last, work of this research wouldn’t have been impossible
without the cooperation of all the members.
10. REFERENCES
[1] Automatic Detection of Melanoma Skin Cancer Using
Texture Analysis, International Journal of Computer
Applications (0975 – 8887) Volume 42– No.20, March
2012. Mariam A.Sheha, Mai S.Mabrouk, Automatic
Detection of Melanoma Skin Cancer Using Texture
Analysis, International Journal of Computer
Applications (0975 – 8887) Volume 42– No.20, March
2012.

[2] Moletest: A Web-based Skin Cancer Screening System,
Dublin Institute of Technology, 2011. ] Jonathan
Blackledge, Moletest: A Web-based Skin Cancer
Screening System, Dublin Institute of Technology,
2011.
[3] A hybrid machine-learning approach for segmentation of
protein localization data, Oxford University Press, Vol.
21 no. 19 2005, pages 3778–3786. Peter M. Kasson1,
Johannes B. Huppa, Mark M. Davis, and Axel T. Brunger,
A hybrid machine-learning approach for segmentation
of protein localization data, Oxford University Press,
Vol. 21 no. 19 2005, pages 3778–3786.
[4] Automatic Diagnosis of Melanoma: a Software System
Based on the 7-Point Check-List, Proceedings of the
43rd Hawaii International Conference on System
Sciences, IEEE, 2010. G. Di Leo, A. Paolillo, P. Sommella,
and G. Fabbrocini, Automatic Diagnosis of Melanoma: a
Software System Based on the 7-Point Check-List,
Proceedings of the 43rd Hawaii International
Conference on System Sciences, IEEE, 2010.
[5] German Capdehourat, Andres Corez, Anabella Bazzano,
and Pablo Mus, Pigmented Skin Lesions Classification
Using Dermatoscopic Images, Springer-Verlag Berlin
Heidelberg, 2009, pp. 537–544
[6] Yogendra Kumar Jain, Comparison between Different
Classification Methods with Application to Skin Cancer,
International Journal of Computer Applications (0975 –
8887) Volume 53– No.11, September 2012
[7] Dr. J. Abdul Jaleel, Artificial Neural Network Based
Detection of Skin Cancer, International Journal of
Advanced Research in Electrical, Elcetronics and
Instrumentation Engineering, Col. 1, Issue 3, 2012
[8] DETECTION OF MELANOMA SKIN CANCER USING
DIGITAL CAMERA IMAGES, ARPN Journal of
Engineering and Applied Sciences, VOL. 10, NO. 7,
APRIL 2015. V. Jeya Ramya, J. Navarajan, R. Prathipa,
and L. Ashok Kumar, DETECTION OF MELANOMA SKIN
CANCER USING DIGITAL CAMERA IMAGES, ARPN
Journal of Engineering and Applied Sciences,

Skin Cancer Detection Application

More Related Content

Skin Cancer Detection Application