Please enable JavaScript.
Coggle requires JavaScript to display documents.
DL for MI (Medical Image) - Coggle Diagram
DL for MI (Medical Image)
Image Type
CT
MRI
XRay
Type : Chest
Name Dataset : Chest X-ray14
Purpose: detection
Number 112.120 images for 30.805 patients
Type : skeletal
Purpose : Detection
Name DataSet : MURA
Ultrasound
PET
Wave images
Biopsy
Mamograph
Purpose : Detection and calssification
Number : 2500 patients
Name Dataset: DDSM
Spectrograms
challenges
Ambiguous boundaries, heterogeneous appearance & varied shape:
complexity of image segmentation
Lack of sufficiently large and labeled image datasets
High model complexity, un- interpretable & hyper parameter optimization
Computationally expensive 3D image segmentation
Network latency and privacy issue
Low resolution images and reconstruction overhead
**
The super-resolution convolutional
neural network is the preferred method for reconstruction
(Umehara et al., 2018).
Solution :
Data Augmentation
Bibliography
Technologies
TensorFlow+keras (Google)
Pytorch (Facebook)
Mindspore (Huawei).
Phases
1: Feature Extraction
2: Dimension Reduction
3: Augmentation
4: Segmentation
5: Clustering/Classification
techniques
Classification
based on combinations
hybrid model
– in this model, the output of convolution
layer is directly passed as input to other DL architecture
such as residual attention network , recurrent convolutional neural
network (RCNN) and inception recurrent residual convolutional
neural network (IRRCNN) model (Alom et al., 2019);
Integrated model
in this model, the output of one DL model
is passed as input to other DL model
Embedded model
– in this
model, the dimension reduction model and classification model
are jointly optimized for example enhanced joint hybrid CNNBiLSTM (EJH-CNN-BiLSTM)
Transfer learning (TL)
–
in this model, DL model trained on one type of problem is used
for the same type of problem Popular CNN models which are
used as TL models are VGG (e.g. VGG16 or VGG19), GoogLeNet (e.g. InceptionV3), Inception Network (Inception-v4), Residual neural Network (e.g. ResNet50), AlexNet.
Joint AB based DL combines
two types of pooling to obtain optimal features: max pooling,
and attentive pooling.
based on training samples
Supervised
long short-term memory (LSTM)
convolutional neural networks (CNNs)
CNN
contains multiple layers which are arranged in a hierarchical fashion. Each layer learns specific features of the
image (Vizcarra et al., 2019). It consists of convolutional layers,
pooling layers, dropout layers, and an output layer.
Architectures
AlexNet
: It consists of 5 convolution and 3 dense layers, max
pooling, dropout, data augmentation, ReLU activations after every
convolutional and fully-connected layer, SGD with momentum
(Krizhevsky et al., 2017).
It is used for object recognition
.
VGG
(Visual Geometry Group): It consists of 13 convolution
layers (in VGG16) & 16 convolution layers (in VGG19), 3 dense layers, pooling, and three ReLU units, very small receptive fields
(Simonyan and Zisserman, 2014).
It is used for large scale object recognition
.
GoogLeNet
. It consisted of 22 layers deep CNN and 4 million
parameters. It contains more filters per layer and stacked convolutional
layers (Zhou et al., 2016).
It used batch normalization, image
distortions, and RMSprop
.
ResNet
(Residual Neural Network): It contains gated units or
gated recurrent units and has a strong similarity to recent successful
elements applied in RNNs.
It is able to train 152 layers NN
(He
et al., 2015).
It has lower complexity than VGGNet
UNet
: It consists of three units: contraction, bottleneck, and
expansion. The contraction section is made of many contraction
blocks. Each block is arranged in a hierarchal fashion. In which
the max-pooling layer is arranged after two convolution layers.
Each block is followed by kernels, whose number is increasing in
multiple of 2. It helps in learning the complex structures. The bottommost layer mediates between the contraction layer and the
expansion layer. It consists of two CNN layers followed by the up
convolution layer.
It performs segmentation and classification in a single step
(Nowling et al., 2019).
DenseNET
Tensor Deep Stack Networks (TDSN)
gated recurrent unit (GRU),
generative adversarial network (
GAN
):
It is used to generate synthetic training data from original
data using latent distribution (Hsieh et al., 2020). It composed of
two networks, a generator, which deployed to generate synthetic
data from noise and a discriminator, which differentiates the real
and synthetic instances of data. Together these two adversarial
networks improve the quality of generated data.
Recurrent Neural Networks (RNNs)
The RNN is very efficient to
capture long term dependencies.
unsupervised
deep belief networks (DBN)
AutoEncoders (AE)
Deep Transfer Network (DTN)
semi-supervised
Segmentation
Format DICOM (Digital
Imaging and Communications in Medicine)