EMS620U Lecture07_Image Segmentation
Processing and Analysis in Medical Imaging
Lecture 7: Image Segmentation (EMS620U) by Stefaan Verbruggen
Image Segmentation: Overview
Image Segmentation Types:
Instance Segmentation
Semantic Segmentation
Panoptic Segmentation
Thresholding
Region-based Segmentation
Edge-based Segmentation
Clustering Segmentation
Medical Image Segmentation
Importance in Medical Imaging:
Critical for annotating and labelling images/videos in datasets.
Extraction of Regions of Interest (ROIs) from various medical images (CT, X-Ray, MRI, US) is fundamental.
Standard image formats include DICOM and NIfTI.
Various segmentation methods exist, including traditional methods and new deep learning techniques.
Instance Segmentation
Definition:
Similar to object detection; e.g., facial recognition.
Detects, labels, and segments every object in the image.
Focuses on delineating object boundaries and identifying overlapping objects.
Useful for tracking individual objects.
Semantic Segmentation
Definition:
Opposite to instance segmentation; provides dense labeling.
Involves classifying every pixel in the image based on intensity (0, 1,...255), transformed into class labels (0, 1,...n).
Effective for segmenting by thresholding.
Panoptic Segmentation
Definition:
Combines instance and semantic segmentation.
Every pixel in the image is labeled, enabling identification of each object.
Provides high granularity in datasets, crucial for detail-oriented tasks.
Thresholding Techniques
Overview of Thresholding:
Simple method dividing pixels into classes based on histogram intensity.
Fixed threshold for low-noise images, but may need dynamic adjustment in noisier ones.
Applicable for single or similarly grouped objects.
Not effective for distinguishing multiple distinct objects.
Challenges in Thresholding
Real-world Issues:
Variabilities in image illumination and device sensitivity can affect segmentation.
Same object may appear with different intensities across an image.
Solution: Adoption of local thresholds.
Approaches to Thresholding
Global Thresholding:
A single value divides greyscale images into background and foreground.
Adaptive Thresholding:
Applies localized thresholds, allowing segmentation to adapt to local image characteristics, improving accuracy.
Local Thresholding Techniques
Block Thresholding:
Divides images into blocks; applies different thresholds for each block—slower but often more effective.
Computing Thresholds
Factors in Threshold Selection:
Conditions vary for different objects; thus, automation of thresholds is often necessary.
Contrast can be analyzed using histograms to guide threshold decisions.
Complex Thresholding Techniques
Multi-Threshold Methods:
Utilizing peak finding for better accuracy despite complexity; trial and error may be required.
Gaussian Mixture Model (GMM):
Otsu’s method fits two Gaussians to maximize variance between pixel classes.
Expectation Maximization (EM) can simultaneously fit multiple classes.
Offers flexibility at the cost of computation power.
Connected Regions in Images
Concepts:
Background vs. Foreground areas; connected regions can represent multiple objects in binary segmentation.
Connected Component Analysis
Purpose:
Allows separate analysis of each detected object in thresholded images (e.g., multiple bones in CT scans).
Assigns distinct labels to non-connected objects based on pixel separation.
Region-based Segmentation
Methodology:
Divides images into regions based on similar features (color, texture, intensity).
Groups pixels until a target segmentation level is reached.
Common techniques include split and merge, often employing AI or manual processes.
Edge-based Segmentation
Technique Overview:
Identifies edges separating objects from backgrounds using criteria changes.
Key Methods:
Canny Edge Detection: Applies Gaussian filter and treats thin edges with hysteresis thresholding.
Sobel Method: Computes gradient magnitude and direction using convolution kernels to identify edges.
Clustering Segmentation
Description:
Groups pixels based on similarity methods; each cluster defines a segment of the image.
Techniques include:
K-means clustering
Mean-shift clustering
Hierarchical clustering
Fuzzy clustering
Summary of Image Segmentation Techniques
Review of segmentation methods:
Instance Segmentation
Semantic Segmentation
Panoptic Segmentation
Thresholding
Region-based Segmentation
Edge-based Segmentation
Clustering Segmentation
Two Common Approaches for Simple Thresholding of an Image
Global Thresholding:
This method uses a single fixed threshold value to separate the pixels of an image into background and foreground. All pixels above the threshold are classified as foreground, while those below are classified as background. This approach works well for images with consistent illumination but may struggle with images that have varying light.
Adaptive Thresholding:
Unlike global thresholding, adaptive thresholding applies localized thresholds to different regions of the image. This allows the segmentation to adapt to local image characteristics, improving accuracy in cases where illumination varies across the image. It analyses smaller sections of the image rather than applying a single threshold value across the entire image.