QUEEN MARY UNIVERSITY
OF LONDON
VIRTUAL MEDICAL IMAGING
LABORTATORY REPORT
EMS620U – Processing and analysis in medical
imaging
, Abstract
The body uses heart scarring as a means of rebuilding damaged heart muscle following a
heart attack. The suppleness and flexibility of this scarred tissue are not comparable to those
of healthy heart muscle. The deep learning approach can be used as better alternative to
manual segmentation to analyse and extract data and useful insights more quickly and
accurately. Deep learning analyses the quantitative data from anomalies, measures rates
and evaluates attributes, and gives data visualisation of the outcomes in relation to blood
flow measurement. Additionally, by using training and testing sets, a machine learning model
can offer improved segmentation. Using machine learning, the training set teaches a
segmentation model to adjust and learn to match the set's data in terms of parameters.
Introduction
Heart scarring is the body’s mechanism in repairing damaged heart muscle after a
heart attack. This scarred tissue does not have the same flexibility and elasticity of healthy
heart muscle. Therefore, there can be complications when the heart is pumping and
transporting blood [1]. Heart scarring can be seen using cardiac MRI imaging. This can be
seen from the delayed washout of gadolinium-based contrast agents [2].
Atrial fibrillation (AFib) is a condition of an irregular heartbeat and a faster than
normal heart rate is observed. A normal heart rate is in the range of 60 – 100 beats per
minute whilst in atrial fibrillation heart rate exceeds 100 beats per minute in some cases.
This can cause feelings of dizziness, shortness of breath and tiredness [3]. The atria are
overloaded with electrical impulses, resulting in atria contracting excessively. This then leads
to blood pooling as the ventricle is not as fast in its contraction compared to atria meaning
that not all the blood is pumped to the blood vessels. Such blood pools can then clot,
eventually leading to stroke and heart failure [4].
Atrial fibrillation occurs due to scar tissue forming in the heart, with the process
known as fibrosis [5]. There have been cases where there is evident heart scarring, however
no symptoms of AFib are present, known as embolic strokes of undetermined source
(ESUS).
Image segmentation is the extraction of regions of interest (ROIs) from medical
images such as Magnetic Resonance Imaging (MRI), Computerised Tomography (CT), X-
Rays and Ultrasound scans [6]. There are multiple types of image segmentation:
1. Instance Segmentation: Deals with detecting and labelling every object in the image
and their boundaries.
2. Semantic Segmentation: Where every pixel is detected and labelled in the image.
3. Panoptic Segmentation: Hybrid method that combines both instance and semantic
segmentation which labels and distinguishes every pixel and labels and distinguishes
individual object instances.
4. Thresholding Segmentation: technique used to classify pixels in segmentation of
grayscale and coloured images [7]
5. Region-based Segmentation: Images are divided into specific regions, using a
method that involves grouping pixels.
6. Edge-based Segmentation: detection of edges of images from a background. Is
used to mark the boundary of an object within an image.
7. Clustering Segmentation: The grouping of pixels based on similarities with each
cluster representing a segment.
OF LONDON
VIRTUAL MEDICAL IMAGING
LABORTATORY REPORT
EMS620U – Processing and analysis in medical
imaging
, Abstract
The body uses heart scarring as a means of rebuilding damaged heart muscle following a
heart attack. The suppleness and flexibility of this scarred tissue are not comparable to those
of healthy heart muscle. The deep learning approach can be used as better alternative to
manual segmentation to analyse and extract data and useful insights more quickly and
accurately. Deep learning analyses the quantitative data from anomalies, measures rates
and evaluates attributes, and gives data visualisation of the outcomes in relation to blood
flow measurement. Additionally, by using training and testing sets, a machine learning model
can offer improved segmentation. Using machine learning, the training set teaches a
segmentation model to adjust and learn to match the set's data in terms of parameters.
Introduction
Heart scarring is the body’s mechanism in repairing damaged heart muscle after a
heart attack. This scarred tissue does not have the same flexibility and elasticity of healthy
heart muscle. Therefore, there can be complications when the heart is pumping and
transporting blood [1]. Heart scarring can be seen using cardiac MRI imaging. This can be
seen from the delayed washout of gadolinium-based contrast agents [2].
Atrial fibrillation (AFib) is a condition of an irregular heartbeat and a faster than
normal heart rate is observed. A normal heart rate is in the range of 60 – 100 beats per
minute whilst in atrial fibrillation heart rate exceeds 100 beats per minute in some cases.
This can cause feelings of dizziness, shortness of breath and tiredness [3]. The atria are
overloaded with electrical impulses, resulting in atria contracting excessively. This then leads
to blood pooling as the ventricle is not as fast in its contraction compared to atria meaning
that not all the blood is pumped to the blood vessels. Such blood pools can then clot,
eventually leading to stroke and heart failure [4].
Atrial fibrillation occurs due to scar tissue forming in the heart, with the process
known as fibrosis [5]. There have been cases where there is evident heart scarring, however
no symptoms of AFib are present, known as embolic strokes of undetermined source
(ESUS).
Image segmentation is the extraction of regions of interest (ROIs) from medical
images such as Magnetic Resonance Imaging (MRI), Computerised Tomography (CT), X-
Rays and Ultrasound scans [6]. There are multiple types of image segmentation:
1. Instance Segmentation: Deals with detecting and labelling every object in the image
and their boundaries.
2. Semantic Segmentation: Where every pixel is detected and labelled in the image.
3. Panoptic Segmentation: Hybrid method that combines both instance and semantic
segmentation which labels and distinguishes every pixel and labels and distinguishes
individual object instances.
4. Thresholding Segmentation: technique used to classify pixels in segmentation of
grayscale and coloured images [7]
5. Region-based Segmentation: Images are divided into specific regions, using a
method that involves grouping pixels.
6. Edge-based Segmentation: detection of edges of images from a background. Is
used to mark the boundary of an object within an image.
7. Clustering Segmentation: The grouping of pixels based on similarities with each
cluster representing a segment.
