Artificial

Intelligence and
Machine Learning
in Medical Imaging
Introduction to AI and ML in
Medical Imaging
Overview of AI and ML: Brief history, definitions, and core concepts.

Role of AI and ML in Medical Imaging: Diagnostic assistance, image

reconstruction, workflow optimization, etc.

Types of ML Models: Supervised, unsupervised, and reinforcement

learning, with examples specific to medical imaging.
 What is Artificial
Intelligence (AI)?

•Definition of AI: “The simulation of human

intelligence in machines designed to think
and learn like humans.”
•Components of AI: Machine Learning (ML),
Deep Learning (DL), Natural Language
Processing (NLP), and Computer Vision.
What is Machine Learning (ML)?

•Definition of ML: “A subset of AI where

machines learn from data and improve
performance over time without explicit
programming.”
•How ML works: Training on data (input-
output relationships), creating algorithms
that make predictions or decisions.
•Examples of ML in daily life: Voice
assistants, recommendation systems,
mobile voice to text and predictive text.
 What is
Machine
Learning (ML)?

Examples of ML in daily
life: Voice assistants,
recommendation systems,
mobile voice to text and
predictive text.
Why AI/ML is Important in
Medical Imaging

•Relevance to Healthcare: AI aids in

handling and interpreting vast amounts
of imaging data, helping address
increasing diagnostic workloads.
•Examples in imaging: Automated
detection of abnormalities, improved
image resolution, reduction of scan
times.
•Benefits for Radiologists and
Patients: Faster diagnostics, improved
accuracy, potential to reduce human
error.
Types of ML Models and Their
Applications in Imaging

•Supervised Learning: Models trained with

labeled data (e.g., classifying images as
“tumor” or “no tumor”).
•Unsupervised Learning: Models find
patterns in unlabeled data (e.g., clustering
MRI scans by abnormality type).
•Reinforcement Learning: Models learn by
trial and error (e.g., optimizing imaging
protocols).
•Examples Specific to Imaging:
Segmenting organs in CT scans
(supervised), clustering diseases in patient
data (unsupervised).
Types of ML Models and Their
Applications in Imaging
Types of ML Models and Their Applications in
Imaging
History and Evolution of AI in Medical Imaging

Early Efforts: Basic computer- Rise of Deep Learning

aided diagnosis (CAD) (2010s): CNNs revolutionized
systems in the 1980s and image-based AI, especially
1990s. with large image datasets.

Present Day: AI integrated in

Future Outlook: Fully
radiology workstations,
automated image analysis,
research into automated
predictive modeling.
diagnosis.
History and Evolution of AI
How Medical Images Are
Processed in AI/ML
Image as Data: Medical images (e.g., X-rays, CT, MRI) are
transformed into pixel data that AI models analyze.

Image Preprocessing: Enhancing image quality, standardizing

formats, and augmenting data for model training.

Labeling and Annotation: Role of radiologists in providing

accurate annotations for supervised learning models.
 Examples of FDA-cleared AI in
radiology

Example of AI automated detection highlighting

critical lung findings on a chest X-ray. Example shown AI anatomical contouring and labelling on an MRI
for Lunit Insight CXR Triage AI cleared by the FDA in rectum exam performed on a Fujifilm radiology
Nov 2021. workstation. AI was cleared in Feb 2021.
Key Terms and Concepts in AI/ML for
Medical Imaging
•Algorithms: The specific set of rules or
calculations that guide decision-making.
•Neural Networks: Computational models inspired
by the human brain, especially useful for image
analysis.
•Training, Validation, and Testing Sets: How data
is split to build reliable models.
•Overfitting vs. Generalization: Balancing model
performance to ensure it works on real-world data.
 Training, Validation, and Testing
Sets
Purpose: To build reliable and accurate machine learning models.
Process:
1.Split Data: Divide your dataset into three subsets:
◦ Training Set: Largest subset for model training.
◦ Validation Set: Used to tune hyperparameters and prevent overfitting.
◦ Testing Set: Used to evaluate the final model's performance.

2.Train the Model: Use the training set to train the model, learning patterns and relationships in the data.
3.Validate the Model: Evaluate the model's performance on the validation set after each training epoch. Adjust hyperparameters and
model architecture as needed.
4.Test the Model: Once training is complete, evaluate the final model's performance on the unseen testing set.
Goal:
◦ Prevent Overfitting: Ensure the model generalizes well to new, unseen data.
◦ Improve Model Performance: Tune hyperparameters and select the best model architecture.
◦ Obtain Unbiased Evaluation: Get an accurate assessment of the model's performance on real-world data.
◦ Common Splits:
◦ 70% Training, 15% Validation, 15% Testing
◦ 80% Training, 10% Validation, 10% Testing
◦ 60% Training, 20% Validation, 20% Testing

Note: The optimal split depends on the size and complexity of your dataset.
Fundamentals of Image
Processing and Analysis
Basic Image Processing Techniques: Filters, edge detection,
segmentation.

Feature Extraction: How images are transformed into data

features suitable for ML algorithms.

Labeling and Annotation: Importance of annotated datasets and

challenges in medical data.
 Basic Image Processing Techniques
•Purpose of Image Processing: Image processing is a technique used to manipulate and analyze digital
images to improve their quality, extract information, and prepare them for further analysis. It involves a
series of algorithms and techniques applied to images to achieve specific goals.
•Techniques:
•Noise Reduction: Removing unwanted noise (artifacts) in images to improve clarity.
•Smoothing and Sharpening: Techniques that improve visibility of edges and structures.
•Contrast Adjustment: Enhancing contrast to distinguish structures more easily, especially in
grayscale images like CT or MRI.
•Image Segmentation: Partition images into meaningful regions or objects. Facilitate analysis and
interpretation.
 What is Image Segmentation?
Image segmentation is a technique used to partition an
image into multiple segments or regions, each representing a
distinct object or area of interest.
Image Why is Image Segmentation Important?
Segmentatio •Object Detection and Recognition: Identify and classify
n objects within an image.
•Medical Image Analysis: Segment organs, tumors, and other
anatomical structures for diagnosis and treatment planning.
 Thresholding:
• Simple technique based on intensity
Common Image values.
• Pixels above a threshold are assigned
Segmentation Techniques:
to one class, and those below to
another.

Image Region-Based Segmentation:

• Groups pixels with similar properties
Edge-Based Segmentation:
• Detects edges or boundaries between

Segmentatio (e.g., color, texture) into regions.

• Region growing and region
regions.
• Can be sensitive to noise and may not

n
splitting/merging are common handle complex objects well.
approaches.

Watershed Segmentation: Machine Learning-Based

• Treats the image as a topographic map Segmentation:
and identifies regions based on • Uses machine learning models (e.g.,
watersheds. convolutional neural networks) to
• Can be sensitive to noise and initial learn complex patterns and segment
seed points. images accurately.
 Noise and Artifacts: Noise can interfere with
segmentation algorithms.

Ill-Defined Boundaries: Objects may not have

Challenges clear edges, making segmentation difficult.
in Image
Segmentatio Varying Illumination: Changes in lighting
n conditions can affect segmentation
performance.
Complex Textures: Textured regions can be
challenging to segment accurately.
 •Medical Image Analysis: Segmenting organs,
tumors, and other anatomical structures for diagnosis
Applications and treatment planning.
•Self-Driving Cars: Segmenting roads, lanes, and
of Image obstacles for autonomous navigation.
Segmentatio •Satellite Image Analysis: Segmenting land use,
vegetation, and water bodies for environmental
n: monitoring.
•Object Detection and Tracking: Segmenting objects
of interest for tracking and analysis.
Feature Extraction: Transforming
Images into Data
What is Feature Extraction?
Feature extraction is the process of extracting meaningful features from raw image data. These
features, which are often numerical values, represent the essential characteristics of the image
that are relevant to a specific task.
Why is Feature Extraction Important?
•Reducing Data Dimensionality: Raw image data can be very high-dimensional. Feature extraction
helps reduce the dimensionality of the data, making it more manageable for machine learning
models.
•Improving Model Performance: By extracting relevant features, we can provide the model with
the most informative information, leading to better performance.
•Enabling Model Interpretation: Feature extraction can help us understand how the model makes
decisions by revealing the underlying patterns in the data.
Common Feature Extraction
Techniques
•Color Features:
•Color Histograms: Quantify the distribution of colors in an image.
•Color Moments: Capture the mean, standard deviation, and skewness of color channels.
•Texture Features:
•Statistical Texture Features: Calculate statistical measures like mean, variance, and entropy of pixel intensities.
•Structural Texture Features: Analyze the spatial arrangement of patterns in an image.
•Shape Features:
•Geometric Features: Measure properties like area, perimeter, and circularity.
•Fourier Descriptors: Represent shapes as a series of coefficients.
•Local Binary Patterns (LBP):
•Capture local texture information by comparing the intensity of a central pixel to its neighbors.
Feature Extraction: Transforming
Images into Data
•Common Features:
•Shape Features: Shape, size, and edges of structures (e.g.,
tumor contour).
•Texture Features: Patterns in pixel intensity that can
indicate tissue density variations.
•Intensity Features: Brightness levels in different regions,
often used for distinguishing tissues.
Application: ML algorithms use these features to make
predictions (e.g., cancerous vs. non-cancerous tissue).
Preprocessing for Machine
Learning

STANDARDIZATION DATA AUGMENTATION ANNOTATION AND

AND NORMALIZATION LABELING
Standardization and Normalization:
Ensuring Consistent Image Intensity Values
Why Standardize and Normalize?
•Handle Variability: Image intensity values can vary due to lighting, camera settings, etc.
•Improve Model Performance: Consistent data leads to better model performance.
•Faster Convergence: Scaled data accelerates model training.
 Standardization and Normalization:
Ensuring Consistent Image Intensity Values
Standardization
• Formula: Z = (X - μ) / σ
• Process:
• Calculate the mean (μ) and standard deviation (σ) of the dataset.
• Subtract the mean from each data point.
• Divide the result by the standard deviation.

Normalization
• Min-Max Normalization: Scales data to a specific range (e.g., 0-1).
• Formula: X_normalized = (X - X_min) / (X_max - X_min)

• Z-Score Normalization: Same as standardization.

• Decimal Scaling: Divides values by a power of 10.
Standardization and Normalization:
Ensuring Consistent Image Intensity Values
Benefits:

Improved model performance

Faster convergence

Better feature comparison

Data Augmentation: Expanding
Your Dataset
What is Data Augmentation?
•Artificially increasing the size and diversity of a dataset.
•Creating modified versions of existing images.
Why is it Important?
•Prevent Overfitting: Reduces the risk of memorizing the training data.
•Improve Generalization: Model learns to recognize patterns under various conditions.
•Reduce Need for More Data: Expands the dataset without additional data collection.
 Data Augmentation: Expanding
Your Dataset
Common Techniques:
• Geometric Transformations:
• Rotation, flipping, scaling, translation, shearing

• Color Transformations:
• Brightness, contrast, color jitter

• Noise Addition:
• Gaussian noise, salt and pepper noise

Considerations:
• Data Quality: Ensure augmented data remains realistic.
• Balance: Maintain class distribution.
• Augmentation Strength: Experiment with different levels.
Example of a medical image
with augmented versions
Annotation and Labeling: The
Foundation of Supervised
Learning
What is Annotation and Labeling?
•The process of assigning labels or tags to data.
•Provides the model with ground truth information.
•Essential for supervised learning algorithms.
Why is it Important?
•Model Training: Labeled data serves as the basis for training machine learning models.
•Performance Evaluation: Labeled data is used to evaluate the model's accuracy and
performance.
•Continuous Improvement: By annotating more data, we can refine and improve the model's
performance.
Examples of Annotation
Image Annotation:
•Object Detection: Bounding boxes around objects of interest.
•Semantic Segmentation: Pixel-level labeling of different objects in an image.
•Instance Segmentation: Identifying and segmenting individual instances of objects.

Text Annotation:
•Named Entity Recognition: Identifying and classifying named entities (e.g., persons, organizations, location
•Sentiment Analysis: Labeling text as positive, negative, or neutral.

Medical Image Annotation:

•Tumor Segmentation: Marking regions of tumor cells in medical images.
•Organ Segmentation: Identifying and labeling different organs in medical scans.
Challenges in Image Processing
for Medical Imaging
Quality and Consistency Issues: Variability in image quality due to
equipment differences or noise.

Large Data Requirements: AI models need large, annotated datasets which

can be resource-intensive to create.

Ethical and Privacy Considerations: Challenges in data sharing and patient

privacy when images are processed and stored for training models.
Overview of Software and Tools
for Image Processing
•Common Tools: MATLAB: Widely used for image analysis in medical research.
•Python Libraries: Such as OpenCV, scikit-image, and TensorFlow.
•DICOM Standards: Format commonly used in medical imaging that includes image metadata.
 Practical Application: Example
Workflow of Image Processing
for AI
1. Image Acquisition:
• Capture images using various modalities (MRI, CT, X-ray, etc.).
• Ensure high-quality images with minimal artifacts.
2. Preprocessing:
• Noise Reduction: Remove unwanted noise or interference from images.
• Contrast Enhancement: Adjust image contrast to improve visibility of features.
• Normalization: Scale image intensities to a common range for consistency.
3. Segmentation:
• Identify Regions of Interest (ROI): Isolate specific areas within the image.
• Techniques: Thresholding, region-based segmentation, edge detection, etc.
 Practical Application: Example
Workflow of Image Processing
for AI
4. Feature Extraction:
• Convert Image Data into Numerical Features: Extract relevant information from images.
• Common Features: Color, texture, shape, and spatial features.
5. Data Augmentation:
• Create Variations of Images: Increase dataset size and diversity.
• Techniques: Rotation, flipping, scaling, cropping, noise addition, color jittering.
6. Model Training:
• Select a Suitable Model: Choose a model architecture (CNN, RNN, etc.).
• Train the Model: Use the augmented dataset to train the model.
• Optimize Hyperparameters: Adjust learning rate, batch size, and other parameters.
Practical Application: Example
Workflow of Image Processing
for AI
7. Model Evaluation:
•Assess Model Performance: Evaluate accuracy, precision, recall, and F1-score.
•Identify Areas for Improvement: Analyze model errors and adjust accordingly.
8. Model Deployment:
•Integrate into Applications: Deploy the trained model into real-world applications.
•Real-time or Batch Processing: Process images in real-time or in batches.
 Preprocessing: Remove
Image Acquisition: Segmentation: Identify
noise, enhance contrast,
Obtain medical images tumor regions or specific
and normalize intensity
like CT scans or MRIs. anatomical structures.
values.

Example:
Feature Extraction:
Extract features like
Data Augmentation:
Create variations of
medical images to
Model Training: Train a
deep learning model to
classify tumors or
Medical
Image
texture, shape, and
improve model predict disease
intensity patterns.
robustness. progression.

Analysis
Model Evaluation: Model Deployment:
Assess the model's Integrate the model into
accuracy on a validation a clinical decision
set. support system.
Key AI/ML Algorithms Used in
Imaging
Convolutional Neural Networks (CNNs): How they work, why
they’re effective for images.

Generative Models: GANs and their role in generating synthetic

images.

Other Algorithms: Decision trees, support vector machines, etc.,

and when they might be applied in imaging.
Overview of Machine Learning
Algorithms
AI and ML: Revolutionizing Medical Image Analysis
Key Benefits:
•Automated Image Analysis: Efficient analysis of large datasets.
•Improved Diagnostic Accuracy: Detect subtle abnormalities.
•Enhanced Decision Making: Support informed clinical decisions.
•Personalized Medicine: Tailor treatment plans to individual patients.
Overview of Machine Learning
Algorithms
•Key Categories of Algorithms:Supervised Learning: Models trained on labeled data.
•Unsupervised Learning: Models that identify patterns in unlabeled data.
•Reinforcement Learning: Models that learn through trial and error to optimize outcomes.
Convolutional Neural Networks
(CNNs)
What is a CNN?
•Definition: A type of deep learning model that is particularly effective for analyzing visual data
(images).
•How It Works: CNNs process images through layers (convolutional, pooling, and fully connected
layers) to extract and learn spatial hierarchies of features.

Key Layers in a CNN:

•Convolutional Layer: Detects edges, textures, and patterns.
•Pooling Layer: Reduces dimensionality, making the model faster and more efficient.
•Fully Connected Layer: Helps in classifying the final image or object.

Application in Imaging: Primarily used for tasks like object detection, classification (e.g., tumor
detection), and segmentation (e.g., organ segmentation in MRI).
Generative Adversarial Networks
(GANs)
What is a GAN?
•Definition: A type of deep learning model that consists of
two neural networks (a generator and a discriminator) that
work against each other to improve performance.
•Generator: Creates fake images or data from random
noise.
•Discriminator: Evaluates images to determine if they are
real or fake.

How GANs Work: The generator tries to fool the

discriminator into thinking its generated images are real,
and through this adversarial process, both networks
improve.
Application in Imaging: Image Synthesis: Generating

Generativ
synthetic medical images for training AI models
(especially useful when annotated data is limited).

e
Data Augmentation: Creating diverse variations of Adversari
existing data to train models more effectively.
al
Networks
Image Enhancement: Improving image quality or
resolution, particularly in modalities like MRI or CT.
(GANs)
 Diagram showing how a GAN works,
example of a real vs. synthetic medical
image.
Normal and abnormal examples of
the Brain Tumor MRI and Br35H-MRI
datasets (the first row) and their
reconstructed images by f-AnoGAN
(the second row) and GANomaly (the
third row) along with their predicted
labels. Labels are marked green if the
prediction matches the true label,
and red if it does not.

https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10043696
Support Vector Machines (SVMs)
What is an SVM?
•Definition: A supervised learning model used for classification and
regression tasks. It creates hyperplanes that best separate classes in the
feature space.
•How It Works: SVM tries to find a hyperplane that maximally separates
data into different classes (e.g., normal tissue vs. tumor tissue).

Key Features:
•Margin Maximization: SVM aims to maximize the margin between
classes to improve generalization.
•Kernel Trick: Allows SVM to work in non-linear spaces by transforming
the data.

Application in Imaging: Classifying abnormal vs. normal areas in images

(e.g., detecting malignant vs. benign tumors in mammograms).
Decision Trees
What is a Decision Tree?
•Definition: A decision tree is a flowchart-like structure where each internal
node represents a decision based on a feature, and each leaf node represents
an outcome (e.g., tumor or no tumor).
•How It Works: The algorithm splits the data at each node based on the best
feature (usually using criteria like Gini impurity or information gain).

Advantages:
•Interpretability: Decision trees are easy to understand and interpret, making
them useful in medical applications where transparency is crucial.
•Non-linear Relationships: They handle non-linear data well.

Application in Imaging: Classifying regions of interest (e.g., distinguishing

between different tissue types in CT scans).
Random Forests
What is a Random Forest?
•Definition: An ensemble learning method that combines multiple decision
trees to improve predictive performance and avoid overfitting.
•How It Works: A random forest builds many decision trees and aggregates
their results to make a final prediction.

Advantages:
•Reduces Overfitting: By averaging predictions from multiple trees, random
forests reduce the risk of overfitting that a single decision tree might
experience.
•Works Well for High-Dimensional Data: Useful for medical images, which
can have many features (e.g., pixel intensities, textures).

Application in Imaging: Used for classifying regions of interest (e.g., detecting

and classifying lesions in mammograms or CT scans).
K-Nearest Neighbours (KNN)
What is KNN?
•Definition: A simple, instance-based algorithm used for classification or
regression. It assigns a class label based on the majority class of its nearest
neighbors.
•How It Works: KNN computes the distance between the test data point and
other points in the dataset, then selects the majority class among the closest
neighbors.

Advantages:
•Simplicity: Easy to understand and implement, often a baseline model for
comparison.
•Works with Small Datasets: Suitable for situations where labeled data is
limited.

Application in Imaging: Classification of new medical images based on

similarity to known labeled images (e.g., categorizing lung X-rays).
Other Relevant Algorithms

Artificial Neural Networks: Basic artificial neural Autoencoders: Used for unsupervised learning, Recurrent Neural Networks (RNNs): Particularly
networks (ANNs) used for simpler image especially in anomaly detection and feature useful for time-series data in imaging (e.g.,
classification tasks. extraction. analyzing medical image sequences).
Applications in Medical Imaging
Detection and Diagnosis: Examples of AI helping radiologists
identify diseases.

Image Reconstruction: AI in enhancing MRI, CT, PET images.

Prognosis and Treatment Planning: AI assisting in predicting

outcomes and planning therapy.
Ethics, Bias, and Challenges
Data Privacy and Security: Patient data concerns in AI systems.

Bias in Medical AI: Impact of biased training data.

Regulation and Validation: Approval processes and ensuring

model reliability.
Future Directions and Research
Trends
Emerging Applications: Potential future use cases in personalized
medicine, remote diagnostics, etc.

AI and Collaborative Tools: Role of AI in multidisciplinary

healthcare.

Trends in Research: Insights into where research is heading.