Deep Learning Approach for Effective Feature Extraction Biomedical Image

Introduction

Medical Imaging Types

Deep Learning Applications in Medical Imaging

What is Deep Learning

Types of Deep Learning Techniques

PCA Networking (PCANet) for Cognitive State Classification of fMRI Images

Artificial Intelligence

Branch of computer science that imitate human intelligence to create machine intelligence

Sensors, robotics & vision

Machine Learning

Development of computer algorithms that can access data and learn themselves.

Deep Learning

Big Data

Data Science

Size: terabytes (TB), petabytes (PB) and exabytes (EB

Blend of various tools, algorithms, and machine learning principles with the goal to discover hidden patterns from the raw data

Development of computer algorithms that can access data and use it (learn the data without human interference to make the prediction)

Learning process : preprocessing, observing and detecting the pattern of data.

Learn from experience>Train machine>Learn from data

Data types

Image, Audio & Video

Numerical – continuous (temperature, salary), discrete (number of students)

Categorical- represent characteristic(red, blue, man, woman)

Time series- sequence of numbers collected at regular intervals over some period of time (number of home sales for 24 months)

Text data

Skills to Apply Machine Learning

Coding skill : data structure, algorithms, OOPS (object oriented programming systems)

Data skill – Preprocessing, extract information, data analysis and visualization.

Math skills – Probability, Statistics, Linear Algebra, calculus

Knowledge of machine learning algorithms -should be familiar unsupervised (clustering, dimensionality reduction) and supervised (regression,decision trees,) techniques

Familiar with ML framework - SCIKIT LEARN, AZURE, TENSORFLOW, CAFFE, THEANO, SPARK & TORCH

Types of Machine Learning

Unsupervised Learning

Reinforcement Learning

Supervised Learning

Classification: Fraud detection, Email Spam Detection, Diagnostics, Image Classification

Regression: Risk Assessment, Score Prediction

Dimensionality Reduction: Text Mining,Face Recognition,Big Data Visualization, Image Recognition

Clustering: Biology, City Planning Targetted Marketing

Reinforcement Learning: Gaming, Finance Sector, Manufacturing, Inventory Management, Robot Navigation.

Types of ML

Supervised Learning Data: (x,y) X is data, y is label Goal: Learn function to map x → y Example: This is a cars

Unsupervised Learning Data: x, x is data, no labels
Goal: Learn underlying Example: Two structure are similar

Reinforcement Learning:
Data: state- action-reward
Goal: Maximize future rewards Example: Learn to drive

Artificial Neural Network (ANN) is based on biological neurons.
They take inputs from other neurons, multiply a weight, add a bias and apply activation function

Machine Learning (Support Vector Machine, PCA, Reinforcement Learning, Clustering, ANN, K-Nearest Neighbor)

Deep Learning (AlexNet, LeNet, ZFNet, GooLENet, ResNet, VGGNet)
PCAnet

Artificial neural networks are composed of units

Units are connected by weights

The output activation about of a unit depends on Incoming activations a0, a1, Weights w0, w1

Multilayer perceptron (MLP)

Feedforward network of artificial neurons, Three types of layers:(Input layer-Contains features for sample, Hidden layer-Combines features, Output layer-Computes output probability for each class)

Convolutional Artificial Network (CNN)

Input Layer

Convolutional Layer

Pooling Layer

Fully Connected Layers

Output Layer

Sliding a set of ‘filters’ or ‘kernels’ across the input data to detect a specific feature or pattern, such as edges, corners etc.

Reduce the spatial size/dimension of the input. (max pooling and average pooling)

Last hidden layer that fully connected to previous layer

Final features learned layer

HOW DEEP LEARNING WORKS?

Change in classifier approach using typical machine learning algorithm and deep learning technique.

Convolutional Neural Network (Feature Extraction and Classification

Input patch size

Convolutional neural networks

Filter kernels & feature maps

Pooling Layer

Separate the image into non-overlapping subimages (e.g. 2x2)

Select the maximum / average / ... in each subimage

Statistics over neighboring features to reduce the size of the feature maps

Ultrasonic Imaging –prenatal imaging, evaluating abdominal and pelvic organs, diagnosing heart conditions, assessing blood flow, guiding needle biopsies, and detecting tumors or cysts in various body regions

Positron Emission Tomography (PET)–
cancer diagnosis, staging, and treatment planning. Assessing brain disorders, heart conditions, and neurological diseases

Computerized Tomography (CT) –
Diagnose and monitor tumors, vascular disease, organ abnormalities, bone fractures

Magnetic Resonance Imaging (MRI)-
detect tumors, stroke, neurological disorders, joint injuries, and provide detailed anatomical information

Learned features technique/ whole image-based - Principle Component Analysis (PCA), Concolutional Neural Network (CNN), PCAnet

PCANet is a PCA-based deep learning technique that is originated from Principle Component Analysis (PCA).

Feature Extraction - Process to extract high-level features or discriminative features from original input data.
Transform raw data into numerical features with smaller size that can be processed without losing information and yields better accuracy.

Manual Interpretation

Interpretation of the diagnosis

CAD System

Reduction of characteristics

Feature classifier

Computer output second opinion

Extraction of characteristics

Pretreatment

Deep learning application

click to edit

Detection : organ/region/remark localization

Classification : Image /object classification

Feature extraction: Extract meaningful features from image

Reconstruction: enhance image quality or generate high-resolution images from low-quality or sparse data

Segmentation: organ/substructure sementation

Application

Techniques

Feature Extraction/Classification

Reconstruction

Segmentation/Detection

Feature Extraction

CNN (AlexNet, LeNet, ZFNet, GooLENet, ResNet, VGGNet), MLP, Restricted Boltz Machine,Deep Boltzman Machine

Variational Autoencoders (VAEs)
Generative Adversarial Networks (GANs)

Convolutional Neural Network (CNN)

PCANet, PCANet ++, PCANet Pooling Raw