Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Sneha Thanvi2
C Durga Devi M.E,1 PG Scholar, Department of Electronics and Communication
Assistant Professor, Department of Electronics and Engineering
Communication Engineering DMI College of Engineering
DMI College of Engineering Palanchur, Chennai, India
Palanchur, Chennai, India
Abstract— Automatic skin lesion segmentation in normal skin without any pre existing eruptions is called a
dermoscopic images is a challenging task due to the low primary lesion.
contrast between lesion and the surrounding skin, the irregular
and fuzzy lesion borders. A fully automated method is used for These include patches, where the only change is color,
skin lesion segmentation by leveraging the discriminative papules, nodules and tumors, which are elevated blisters,
power of a 19 layer deep convolutional neural network. Inorder cysts and pustules, which contain serum, keratinized
to update the result of convolutional neural network, fully substances, pus etc and urticaria which is temporarily
convolutional network is implemented. Fully convolutional elevated. In the case of malignant lesions, i.e., skin cancer,
network is used for automatically detecting efficient output at the cells split quickly, and may invade other parts of the
any situation. A loss function is based on Jaccard distance is body. Indeed, these cells do not die, as generally occurs with
used to eliminate the need of sample re-weighting. Jaccard normal cells. Types of primary skin lesions are:
distance is the type of loss function which is used as strong
imbalance between the foreground and background pixels. Papule:
Fully convolutional network need minimum pre and post A papule is a localized elevated lesion of 10 mm or less
processing, which allow its adoption in variety of medical
in diameter with a hemispheric or flat shape. It is
image segmentation task.
characterized by a surface that can be smooth, eroded,
ulcerative, hyperkeratosis or crusted. It may be caused by a
Keywords— loss function, Convolutional Neural Network
proliferative or inflammatory change in the epidermis, or by
I. INTRODUCTION dermal edema.
Leukoderma:
A skin lesion is a part of the skin that has an abnormal
growth or appearance compared to the skin around it. A skin Leukoderma is a white patch produced by
lesion is generally called an eruption. Eruptions are divided Depigmentation or local anemia. Depigmentation is caused
into primary lesions, which occur in normal skin, and by abnormal production of melanin’s such as in vitiligo
secondary lesions, which are caused secondarily by other vulgarisms. Nevus anemic us causes local anemia leading to
eruptions. leukoderma. Leukoderma in the periphery of an eruption is
called a white halo.
Nodule and Tumor:
A nodule is a localized lesion that appears as a papule
with a diameter of 10 to 20 mm. It can have various causes,
such as tumor formation, granulomatous change,
inflammation or edema. An intensely proliferative nodule
with an elevation of 30 mm or more in diameter is called a
Fig 1.1 Skin Lesion tumor.
Computational systems for skin cancer diagnosis have Blister:
been proposed in order to aid dermatologists in early
assessment of such diseases, or even to monitor pigmented A blister is a skin elevation of 5 mm or more in diameter
skin lesions. Dermoscopy images are widely applied for enclosed by a membrane and containing transparent fluid that
image analysis of pigmented skin lesions. Such images may is mainly plasma and cellular material. A small blister with a
be acquired from dermatoscope devices or specific cameras diameter of less than 5 mm is called a vesicle. A hemorrhagic
to provide a better visualization of the pigmentation pattern blister containing serum mixed with blood is referred to as a
on the skin surface. bloody bulla.
Primary Skin Lesion are abnormal condition present at Pustule:
birth or over a person’s lifetime. An eruption that occurs in A pustule is a yellowish blister with purulent contents
III. Background of the Study Hoo-Chang Shin, Member, Holger R. Roth, Mingchen
Gao, Le Lu, Ziyue Xu, Isabella Nogues, Jianhua Yao, Daniel
Mollura, Ronald M. Summers proposed a method based on
Harald Ganster ,Axel Pinz, Reinhard Röhrer, Ernst deep CNN. Remarkable progress has been made in image
Wildling, Michael Binder, and Harald Kittler proposed the recognition, primarily due to the availability of large-scale
automated method for the recognition of melanoma. A annotated datasets and the revival of deep CNN. CNNs
system for the computerized analysis of images obtained enable learning data-driven, highly representative, layered
from ELM has been developed to enhance the early hierarchical image features from sufficient training data.
recognition of malignant melanoma. As an initial step, the However, obtaining datasets as comprehensivelyannotated as
binary mask of the skin lesion is determined by several basic ImageNet in the medical imaging domain remains a
segmentationn algorithms together with a fusion strategy. A challenge. There are currently three major techniques that
set of features containing shape and radiometric features as successfully employ CNNs to medical image classification
well as local and global parameters is calculated to describe training the CNN from scratch, using off-the-shelf pretrained
the malignancy of a lesion. Significant features are then CNN features, and conducting unsupervised CNN pre-
selected from this set by application of statistical feature training with supervised fine-tuning. Another effective
subset selection methods. The final KNN classification method is transfer learning, i.e., fine tuning CNN models.
delivers a sensitivity of 87% with a specificity of 92%. a Two specific CAD problems, namely thoraco-abdominal
fully connected feed-forward network with 21 input units lymph node detection and ILD classification are used. This
that represent 21 skin lesion features one hidden layer with method achieves the state-of-the-art performance on the
four hidden units, and one output unit for the categorization mediastinal LN detection, with 85% sensitivity at 3 false
into benign or malignant class. After training and application positive per patient, and report the first five-fold cross-
validation classification results on predicting axial CT slices method was validated in the Brain Tumor Segmentation
with ILD categories. Extensive empirical evaluation, CNN Challenge 2013 database obtaining simultaneously the first
model analysis and valuable insights can be extended to the position for the complete, core, and enhancing regions in
design of high performance CAD systems for other medical Dice Similarity coefficient metric for the Challenge data set.
imaging tasks. Also, it obtained the overall first position by the online
evaluation platform.
Lequan Yu, Student Member, Hao Chen, Qi Dou, Jing
Qin, and Pheng-Ann Heng proposed a method for melanoma Shun Miao, Member, Z. Jane Wang, Senior Member, and
recognition using residual networks. Automated melanoma Rui Liao, proposed a CNN regression approach to address
recognition in dermoscopy images is a very challenging task the two major limitations of existing intensity-based 2-D/3-D
due to the low contrast of skin lesions, the huge intra class registration technology: 1) slow computation and 2) small
variation of melanomas, the high degree of visual similarity capture range. Different from optimization-based methods,
between melanoma and non-melanoma lesions, and the which iteratively optimize the transformation parameters
existence of many artifacts in the image. In order to meet over a scalar-valued metric function representing the quality
these challenges, a novel method for melanoma recognition of the registration, the proposed method exploits the
by leveraging very deep CNNs is used. Compared with information embedded in the appearances of the digitally
existing methods employing either low-level hand-crafted reconstructed radiograph and Xray images, and employs
features or CNNs with shallower architectures, substantially CNN regressors to directly estimatethe transformation
deeper networks can acquire richer and more discriminative parameters. An automatic feature extraction step is
features for more accurate recognition. To take full advantage introduced to calculate 3-D pose-indexed features that are
of very deep networks, a set of schemes is used to ensure sensitive to the variables to be regressed while robust to other
effective training and learning under limited training data. factors. The CNN regressors are then trained for local zones
First, the residual learning is applied to cope with the and applied in a hierarchical manner to break down the
degradation and overfitting problems when a network goes complex regression task into multiple simpler sub-tasks that
deeper. This technique can ensure that the networks benefit can be learned separately. There are currently three major
from the performance gains achieved by increasing network techniques that successfully employ CNNs to medical image
depth. Then, a FCRN is constructed for accurate skin lesion classification: training the “CNN from scratch” ,using “off-
segmentation, and further enhance its capability by the shelf CNN” features as complementary information
incorporating a multi-scale contextual information channels to exist hand crafted image features, performing
integration scheme. Finally, seamlessly integrate the unsupervised pre training on natural or medical images and
proposed FCRN and other very deep residual networks to fine-tuning on medical target images using CNN or other
form a two-stage framework. This framework enables the types of deep learning models. Weight sharing is further
classification network to extract more representative and more employed in the CNN regression model to reduce the
specific features based on segmented results instead of the memory footprint. This approach has been quantitatively
whole dermoscopy images, further alleviating the evaluated on three potential clinical applications,
insufficiency of training data. This framework is extensively demonstrating its significantly advantage in providing highly
evaluated on ISBI 2016 Skin Lesion Analysis Towards accurate real-time 2-D/3-D registration with a significantly
Melanoma Detection Challenge dataset. Experimental results enlarged capture range when compared to intensity based
demonstrate the significant performance gains of the methods.
proposed framework, ranking the first in classificationand the
second in segmentation among 25 teams and 28 teams, Qi Dou, Hao Chen, Lequan Yu, Lei Zhao, Jing Qin,
Defeng Wang, Vincent CT Mok, Lin Shi and Pheng-Ann
respectively. This method corroborates that very deep CNNs
with effective training mechanisms can be employed to solve Heng proposed a novel automatic method to detect CMBs
from magnetic resonance images by exploiting the 3D CNN.
complicated medical image analysis tasks, even withlimited
training data. CMBs are small haemorrhages nearby blood vessels. They
are recognized as important diagnostic biomarkers for many
Sergio Pereira, Adriano Pinto, Victor Alves, and Carlos A. cerebro vascular diseases and cognitive dysfunctions. In
Silva used a CNN method for the segmentation of brain current clinical routine, CMBs are manually labelled by
tumors. Among brain tumors, gliomas are the most common radiologists but this procedure is laborious, time consuming
and aggressive, leading to a very short life expectancy in and error prone. Compared with previous methods that
their highest grade. Thus, treatment planning is a key stage to employed either low-level hand-crafted descriptors or 2D
improve the quality of life of oncological patients. MRI is a CNNs, this method can take full advantage of spatial
widely used imaging technique to assess these tumors, but contextual information in MR volumes to extract more
the large amount of data produced by MRI prevents manual representative high-level features for CMBs, and hence
segmentation in a reasonable time, limiting the use of precise achieve a much better detection accuracy. To further improve
quantitative measurements in the clinical practice. So, the detection performance while reducing the computational
automatic and reliable segmentation methods are required; cost, a cascaded framework is used under 3D CNNs for the
however, the large spatial and structural variability among task of CMB detection. First, exploit a 3D FCN strategy to
brain tumors make automatic segmentation a challenging retrieve the candidates with high probabilities of being
problem. In this method, an automatic segmentation method CMBs, and then apply a well trained 3D CNN discrimination
is used based on CNN exploring small 3*3 kernels. The use model to distinguish CMBs from hard mimics.Compared
of small kernels allows designing a deeper architecture, with traditional sliding window strategy, the proposed 3D
besides having a positive effect against overfitting, given the FCN strategy can remove massive redundant computations
fewer number of weights in the network. This method makes and dramatically speed up the detection process. A large
the use of intensity normalization as a pre-processing step, dataset is applied with 320 volumetric MR scansand
which though not common in CNN based segmentation performed extensive experiments to validate this method,
methods, proved together with data augmentation to be very which achieved a high sensitivity of 93.16% with an average
effective for brain tumor segmentation in MRI images. This number of 2.74 false positives per subject, outperforming
previous methods using low-level descriptors or 2D CNNs The figure 3.1 shows the block diagram of
by a significant margin. This method, in principle, can be Methodology
adapted to other biomarker detection tasks from volumetric
medical data.
Zhennan Yan, Yiqiang Zhan, Zhigang Peng, Shu Liao,
Yoshihisa Shinagawa, Shaoting Zhang, Dimitris N. Metaxas,
and Xiang Sean Zhou designed a multi-stage deep learning
framework for image classification and apply it on body part
recognition. In general image recognition problems,
discriminative information often lies in local image patches.
For example, most human identity information exists in the
image patches containing human faces. The same situation
stays in medical images as well. “Body part identity” of a
transversal slice which body part the slice comes from - is
often indicated by local image information, e.g. a cardiac
slice and an aorta arch slice are only differentiated by the Fig. 3.1.Block diagram of methodology
mediastinum region. Specifically, this method aims at: 1)
discover the local regions that are discriminative and non-
3.2 FCN ARCHITECTURE
informative to the image classification problem, and 2) learn
a image-level classifier based on these local regions. These Pixel-wise classification is performed and FCN is
two tasks are achieved by the two stages of learning scheme, essentially served as a filter that projects the entire input
respectively. In the pre-trained stage, a CNN is learned in a image to a map where each element represents the
multi instance learning fashion to extract the most probability that the corresponding input pixel belongs to the
discriminative and non-informative local patches from the tumor.
training slices. In the boosting stage, the pre-learned CNN is
further boosted by these local patches for image There are two path used in FCN layer. They are,
classification. The CNN learned by exploiting the Convolutional path
discriminative local appearances becomes more accurate than Deconvolutional path
those learned from global image context. The key hallmark
of this method is that it automatically discovers the In contrast image segmentation calls for classifying
discriminative and non-informative local patches through each pixel into foreground or background in combination
multi instance deep learning. Thus, no manual annotation is with full resolution output. Figure 3.2 shows the FCN
required. This method is validated on a synthetic dataset and Architecture
a large scale CT dataset. It achieves better performances than
state-of-the-art approaches, including the standard deep
CNN.
Adriano Pinto, Sergio Pereira, Higino Correia, J.
Oliveira, Deolinda M. L. D. Rasteiro and Carlos A. Silva
proposed a discriminative and fully automatic method for the
segmentation of gliomas, using appearance- and context-
based features to feed an Extremely Randomized Forest
Gliomas are among the most common and aggressive brain
tumours. Segmentation of these tumors is important for
surgery and treatment planning, but also for follow-up
evaluations. However, it is a difficult task, given that its size Fig 3.2 FCN Architecture
and locations are variable, and the delineation of all tumor
tissue is not trivial, even with all the different modalities of With the successive convolution and pooling
the MRI. Some of these features in this method are computed layers, CNNs can integrate contextual information from
over a non-linear transformation of the image. The method regional to global scales, resulting in reduced resolution in
was evaluated using the publicly available Challenge the output layer. In order to address this conflict between
database from BRATS 2013, having obtained a Dice score of multi-scale information aggregation and full-resolution pixel
0.83, 0.78 and 0.73 for the complete tumor, and the core and wise classification, a strategy of up-sampling and de-
the enhanced regions, respectively. The results are convolutional layers is used to recover lost resolution while
competitive, when compared against other results reported
carrying over the global perspective from pooling layers.
using the same database.
The up-sampling layer performs the reverse
IV. PROPOSED SYSTEM operation of pooling and reconstructs the original size of
activation and the de-convolutional layer densifies the
coarse activation map obtained from up-sampling through
3.1 INTRODUCTION swapping the forward and backward passes of a
A fully automatic framework based on deep convolution, thus a single input activation is projected into
convolutional neural network is used for skin lesion multiple outputs after deconvolution, yielding an enlarged
segmentation on dermoscopic images. and dense feature map.
Echo to the convolution path where image
information is aggregated from fine details to global
concept, a hierarchical structure of deconvolutional layers is
used to recover image details at different levels, with the
lower layers encoding overall image information and higher
Equation (3.3) is used to calculate the loss function
layers capturing fine details regarding skin tumors. In this
for each iteration during optimization. With this loss
way, the network can take both global information and fine
function, a weight map is not needed to rebalance pixels
details into account for tumor segmentation.
from tumor region and background. Meanwhile, the
proposed loss function is differentiable which can be
3.3 LOSS FUNCTION
efficiently integrated into back propagation during network
training.
For any pixel in the input image, the corresponding
FCN output represents an estimated posterior probability
3.4 FCN TRAINING
that this pixel belongs to the skin tumor the skin tumor.
Training a deep network model with a limited
Since FCN essentially performs a pixel-wise classification,
number of samples is a challenging task. FCN has 19 layers
cross entropy is usually used as the loss function.
and 290, 129 parameters to be learned. Even though the
number of training samples available is considered as a
large-scale study in the field of digital dermoscopic image
analysis, it is still relatively small as compared to the size of
Where, ∈{0, 1} is the actual class of with = 1 for tumor and the network. In order to address this issue, firstly apply FCN
= 0 for background. Using equation the loss function is model to convert image segmentation into a pixel-wise
calculated for cross entropy. classification problem. As the result, each pixel can be
In Dermoscopy, a skin tumor usually occupies a considered as an independent sample during network
small region in the whole dermoscopic image. As the result, training, which greatly boosts the number of equivalent
the pixel-wise classification tends to be biased towards the training samples. Then, initialize the network weights using
background and therefore there is a higher chance for tumor Xavier’s technique and employ the following strategies to
to be partially segmented or even missed. A typical solution improve the efficiency of network training while reducing
is to assign a weight to each pixel during training that overfitting.
compensates the different frequencies of pixels from each
class, so the contributions of tumor and background are re- 1) Adam stochastic optimization:
balanced. However, this pixel wise reweighting procedure Stochastic gradient descent (SGD) with mini-batch
brings additional computation cost especially when image is usually employed as the optimization algorithm for neural
augmentations are employed in the training procedure. network training. It is well known that learning rate is one of
Novel loss function based on Jaccard distance. Jaccard index the critical hyper parameters that have a significant impact
is used, which is also known as the Jaccard similarity on classification performance. However, selecting proper
coefficient, is one of the most frequently used evaluation learning rate and strategy can be fairly challenging.
measures in medical image segmentation.
Then Adam optimization algorithm, or adaptive
The Jaccard distance, which measures dissimilarity moments, is adapted to adjust the learning rate based on the
between two sets, is complementary to the Jaccard index. first and the second-order moments of the gradient at each
Let M represent the ground truth of segmentation, which is iteration.
normally a manually identified tumor region, and C Here the momentum is incorporated as an estimate
represent a computer-generated mask, the Jaccard distance is of the first moment and the effective step size at each
defined as iteration depends on the ratio between the biascorrected first
and second-order moments with a smaller value indicating
that there is greater uncertainty about whether the direction
of corresponds to the direction of the true gradient. Adam is
fairly robust to the choice of hyper parameters, and set the
learning rate as 0.003 to speed up the training procedure in
where, this method.
M represent the ground truth of segmentation,
which is normally a manually identified tumor region. 2) Batch normalization:
Batch normalization is employed to reduce the
C represent a computer-generated mask internal covariate shift by normalizing the input distribution
of every layer to the standard Gaussian distribution for each
dJ (M,C) itself is not differentiable, which makes it difficult training mini-batch. For this purpose, add a batch
to be directly applied into back propagation. Using equation normalization layer to the output of every convolutional and
(3.2) the Jaccard distance is calculated. Even some de-convolutional layer. The batch normalization is
minimization methods for non differentiable functions can performed over all locations in the same feature map such
be used, but it would be computationally expensive to that different elements in the same map are normalized in
generate a binary mask from continuous FCN output for the same way. In this method, 18 is taken as the batch size.
each iteration during optimization. Thus, it uses the
following loss function 3) Dropout:
Dropout provides a powerful but computationally the total number of epochs is set as 500, and the optimal
inexpensive way to reduce overfitting when training a very epoch that yields the best performance on the validation
deep FCN model with limited data. This technique sets the dataset is saved. Also saved is the corresponding trained
output of each neuron in a given layer to zero with FCN model. After cross validation is completed, another
probability p, thus removes the contribution of those FCN model is trained using the entire training data and the
dropped out neurons from both forward pass and back- total number of epochs is set as the maximum value among
propagation. The subset of disabled neurons is drawn the optimal epochs obtained from cross validation. The
independently for each mini batch and forms a different outputs of these six FCN models are then averaged to
network architecture, then dropout trains the ensemble of all predict the final segmentation.
sub-networks that have different architectures but share
weights in one epoch. In this way, a neuron cannot rely on
the presence of particular other neurons and it is therefore,
forced to learn more robust features that are useful among V. RESULT
different random subsets. This makes the trained FCN model 4.1 TRAINED IMAGES
more robust and improves the generalization ability. The Input dataset image is trained. The figure 4.1
shows the different types of trained lesion images taken
4) Image augmentation: from the dataset.
In order to improve the robustness of the proposed
FCN model under a wide variety of image acquisition
conditions, image augmentation is used to further reduce
model overfitting by artificially enlarging the training
dataset with various image transformations. For the tumor
segmentation on dermoscopic images, primarily look for
invariance to mild geometric transformation as well as
robustness to pixel value variations, so the following two
types of image augmentations is implemented. These
augmentations only require little extra computation, so the
transformed images are generated from the original images
for every mini-batch within each iteration. The other type of
image augmentation focuses on randomly normalizing the
contrast of each channel in the training images.
Fig 4.1 Trained images
3.5 IMPLEMENTATION
Training data fit a model. Unsupervised learning refers to
FCN-based lesion segmentation algorithm was
analysis in which one attempts to learn something about the
implemented with Python based on Theano and Lasagne
data. other than predicting an output value of interest.
packages. The aim is to have the FCN model automatically
learn those features that are useful for lesion segmentation,
so a simple pre-processing is employed to facilitate the
4.2 TRAINED LABELS
following learning procedure while preserving the original
By using thresholding the training images are
image information. By observing most of images in the
converted in to training labels. The figure 4.2 shows the
training set have a height to width ratio of 3:4, resize the
trained labels of lesion image.
images to 192×256 using bi-linear interpolation. RGB
channels are kept as the input to the FCN model and each
channel is rescaled to [0, 1]. The output of FCN model is a
posterior probability map where each pixel value represents
the probability that the pixel belongs to the lesion. A dual-
threshold method is used to obtain a binary tumor mask
from this probability map. Specifically, a relatively high
threshold (thH= 0.8) is firstly applied to the FCN output to
generate a series of tumor candidates. The mass of each
candidate region, which is the area weighted by the pixel
values within this region, is calculated. The tumor center is
chosen as the centroid of the region that has the largest mass
among these candidates. Then a lower threshold = 0.5) is
applied to the probability map. After filling small holes with
morphological dilation, the final tumor mask is determined
as the region that embraces the tumor center. Since this post- Fig 4.2 Trained Labels
processing only involves thresholding and morphological
operations, the whole lesion segmentation can be done Labeled data is a group of samples that have been
within two seconds. tagged with one or more labels. Labeling typically takes a
Finally, a bagging-type ensemble strategy is set of unlabeled data and augments each piece of that
implemented to combine outputs of different FCNs to unlabeled data with meaningful tags that are informative.
further improve the image segmentation performance on the
testing images. For each fold in the 5-fold cross validation,
After obtaining a labeled dataset, the models can be 4.5 SEGMENTED CONTOUR
applied to the data so that new unlabeled data can be The result of image segmentation is a set of
presented to the model and a likely label can be guessed or segments that collectively cover the entire image or a set of
predicted for that piece of unlabeled data. contours extracted from the image.
Relu layer will apply an elementwise activation [2] C. A. Barker and M. A. Postow, “Combinations of
function. A unit employing the rectifier is also called a radiation therapy and immunotherapy for melanoma: a
rectified linear unit. The function returns 0 if it receives any review of clinical outcomes,” Int J Radiation Oncol Biol
negative input, but for any positive value it returns that Phys, vol. 88, no. 5, pp. 986–997, 2014.
value back. [3] C. M. Balch, J. E. Gershenwald, and S.-J. S. et al., “Final
version of 2009 AJCC melanoma staging and classification,”
J Clin Oncology, vol. 27, no. 36, pp. 6199–6206, 2009.
4.8 SEGMENTED OUTPUT IMAGE
After obtaining the features, the image is [4] M. Vestergaard, P. Macaskill, P. Holt, and S. Menzies,
segmented to detect the portion of lesion. The fig 4.8 shows “Dermoscopy compared with naked eye examination for the
the image of segmented output. diagnosis of primary melanoma: a meta-analysis of studies
performed in a clinical setting,” Br. J. Dermatol., vol. 159,
no. 3, pp. 669–676, 2008.