Skin Lesion Segmentation Using Deep Learning Algorithm and Level Set Method
No Thumbnail Available
Date
2019-06-28
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Skin lesion can be benign or skin cancer. Skin cancer is one of the most dangerous cancers
killing so many people all over the world. Skin cancer is the most curable cancer if skin lesions
are diagnosed at early stage. Skin lesion segmentation is a crucial phase in automated skin lesion
detection towards skin cancer. Segmentation of skin lesion is the most challenging task for
dermatologists. The rest phases of computer analysis diagnosis of skin cancer mainly depend on
the segmentation result. Due to this, many methods of skin lesion segmentation have been
proposed to improve the segmentation technique performance in computer aided diagnosis.
In this work, skin lesion segmentation using convolution de-convolution neural network and
contour level set method is used to segment dermoscopic skin lesion images. Convolution deconvolution
neural
network
is
trained
pixel
wise
for
semantic
segmentation
of
pixels
into
lesion
and
background. Level set is used to find the exact edges of detected lesion boundary by
convolution de-convolution neural network method. In addition to the two main proposed
techniques, preprocessing of the input images is applied to remove unwanted artifacts such as
hair over the skin lesion image using vector filters and data augmentation to overcome the over
fitting problem of proposed deep learning network. 2017 International Skin Imaging
Collaboration (ISIC) archive dataset hosted by International Society of Biomedical Imaging
(ISBI) for skin lesion analysis towards melanoma detection is used.
The performance evaluations on the proposed skin lesion segmentation method is pixel wise
average measurements validated against ground truth for test data set are 94.8% intersection over
union, 98.80% specificity, 94.84% sensitivity, 97.84% positive predicted value and 95.58%
negative predicted value. The proposed method out performs segmentation using convolution deconvolution
neural network and level set method by more than 2% and 30% respectively.
Therefore, using convolution de-convolution neural network with level set segmentation method
of skin lesion results better than convolution de-convolution neural network segmentation and
level set segmentation.
Description
Keywords
Dermoscope, Detection, Skin Lesion, Skin Cancer, Segmentation, CDNN, Level Set