This document describes a system for using deep learning and smartphone cameras to detect skin cancer. It trains an Inception neural network on skin lesion images to classify them as malignant or benign. The system is designed to run on low-cost hardware like Raspberry Pi to make early skin cancer detection more accessible. It achieved 76% accuracy on a test dataset. Future work aims to improve this through more complex models, domain knowledge, and generating additional training data.
1. Lightweight Deep Learning on Smart
Device for Early Detection of Skin
Cancer
Dantong Yu
Associate Professor
School of Management
New Jersey Institute Of Technology
2. What is Skin Cancer ?
• Uncontrolled growth of abnormal skin cells
• Often caused by ultraviolet radiation from sunshine or
tanning beds
• Potential Genetic basis for susceptibility
• A major public health problem, with
over 5 million newly diagnosed cases in
the United States each year.
3. How do physicians diagnose skin cancer?
•A skin examination by a dermatologist is the way to
get a definitive diagnosis of skin cancer.
•In many cases, the appearance alone is sufficient to
make the diagnosis.
•Melanoma is the deadliest form of skin cancer,
responsible for over 9,000 deaths each year.
5. Smartphone and IoT Based Assistant
We propose a smartphone and IoT devices based skin cancer
detection system that utilizes deep learning and low-cost
camera to take the snapshots of suspected skin lesions and
distinguish between malignant and benign melanoma skin
images.
6. Related Work
• R. A. Novoa et. al., Dermatologist-level classification of skin
cancer with deep neural networks. Nature 542 (Feb. 2017).
• Y. Li, A. Esteva*, R. Novoa, J. Ko, S. Thrun, Skin Cancer
Detection and Tracking using Data Synthesis and Deep
Learning,
NIPS Machine Learning for Healthcare Workshop 2016
• Apple iTune and Google play SkinVision
– Prevent, Detect and Track Skin Cancer
7. Deep Learning Background
• In our method we use a variant of deep learning models
called Convolutional Neural Networks (CNN)
• A typical CNN comprises of Convolutional, Pooling and Fully
Connected layers
8. Color (RGB) image
One 4x4x3 filter (cube)
The result of a single filter
Another filter
9. A first layer of filters
A second layer of
filters
An MNIST image
A third layer of filters
(treat this as 1176-D
vector)
A “fully connected”
layer
11. Challenges
• Training a Deep learning model from scratch is a time and
resource consuming task. Not sufficient labeled data.
• To ameliorate this problem we adopt a Transfer Learning
Based approach
12. Inception Net
• We adopt Inception Net, a deep Learning model trained on ImageNet
dataset to classify natural images as our base model
• Training a model such as Inception Net can take even weeks on a high
end GPU
• Therefore we fine-tuned Inception Net for the task of skin lesion
classification into Malignant and Benign Melanoma
Inception Net
Architecture
14. Raspberry PI
• The Raspberry Pi 3 is single-board, low end computing device which
has a quad-core Cortex-A53 processor with 1 GB RAM
• Can support camera as well as display devices
• Costs around $35
• Supports Linux OS
15. Training deep learning model on Raspberry PI
• Train the model in a cluster of
Raspberry PI in a distributed manner
• Used Tensorflow framework for
training
• Between-graph replication and
synchronous training for parameter
updates.
16. Results
• Experiment done on the dataset
provided by the ISIC for the 2017
challenge on melanoma detection
and classification.
• After the Transfer Learning step we
were able to achieve an accuracy of
76% on a dataset of 2750 images
with 521 malignant and 2229 benign
images
17. Advantages of our approach
• Makes the process cheap and readily available to masses
• No network connectivity required therefore can be deployed
to remote and diverse regions without any problems
• Can be re-trained as per requirement on a low end device
18. Movidius for speed up
• Movidius stick an another low
end device specifically designed
to speed up inference using a
deep learning model
• In our experiments speed up of
about 5 times was observed
19. Future Works
• Training complex networks to improve
accuracy.
• Integrate domain-knowledge, such as
combining the “ABCDE” characteristics
(Asymmetry , Border irregularity,
Color, Diameter, Evolving size), SPIE
2018 Medical Image
• Enrich Dataset with GAN and InfoGan
by embedding cancer sample into body
and creating more realistic looking
images and evolution history for
training advanced network models.