.
Imam Abdulrahman Bin Faisal University
College of Engineering
Department of Biomedical Engineering
BIOIN 573
Magnetic Resonance Imaging
Final Project
Academic year: 5th year
Submitted by:
Noor Wail Tafish 2200009071
Raghad Aljohani 2200001922
Fatimah Alnahdi 2200004600
Maryam Alzuabi 2200002292
Course Instructor: Dr. Gameel Saleh
Submission Date: 4 May 2024
Abstract
Cancer is considered a life-threatening disease worldwide. Although the existence of various
clinical approaches for cancer detection, Magnetic Resonance Imaging (MRI) has known as a
powerful modality for the detection and characterization of cancer, with its ability to generate high-
resolution images and provide accurate information about tumor size and shape.
The aim of this paper is to provide a comprehensive overview of MRI-based cancer detection,
including important aspects such as the fundamentals of MRI, preprocessing techniques,
segmentation of cancer tissues, machine learning approaches for cancer detection, and finally the
clinical applications and challenges associated with this field.
1. Introduction
Cancar is considered one of the most common diseases currently, and one of the main causes of
death globally. Cancer is abnormal cells that spread and growth without any control, which may
lead to cell aggregation and then tumor formation. The tumor could be benign or malignant tumor,
which are two different types of tumors that differ in the speed of growth and symptoms, as well
as the treatment method. In fact, benign tumor is not cancer since it grows slowly, and it does not
spread. Also, not all the benign tumors need treatment. On the other hand, Malignant tumor can
spread and grow quickly. In addition, it is considered as a cancerous tumor [1].
To identify and detect the presence of tumor or cancer there are many methods that aim to
determine the type and extent of it. Imaging techniques are mostly used to diagnose and detect the
cancer, which allows the observation of internal body structures. These techniques include X-ray,
Computed Tomography (CT) Scan, Magnetic Resonance Imaging (MRI), Ultrasound, and Positron
Emission Tomography (PET) Scan. Each technique has its uses and features that make it eligible
for use as needed. Also, Biopsy is used to differentiate between the cancerous and normal cells,
which includes taking a sample from the tissue and examining this sample under the microscope.
Biopsy could be needle or surgical according to the place and amount of needed sample. In
addition, some types of cancers could be detected using blood tests, which are performed to find
some substances that could be found in the blood during the presence of cancer [2].
In this project the focus will be on one of diagnostic techniques, which is Magnetic Resonance
Imaging (MRI). MRI could help in tumor detection and characterizing with high-resolution images
and accurate identification of the size and shape of tumor [3].
2. MRI Fundamentals
The basic principle of MRI depends on a small magnet moment produced by charged particles in
motion. For example, nuclei-charged particles that contain an odd number of protons have a
characteristic motion called precession that produces a small magnetic moment. When the human
body is positioned in a large static magnetic field (0.5-1.5 Tesla), many of the hydrogen nuclei
align themselves in the same direction of the magnetic field and precess around it. This behavior
is called Larmor precession and is expressed as ω0 = γ B0 , where ω0 is the Larmor frequency, γ
is the gyromagnetic ratio, and B0 is the static magnetic field. After that, a radiofrequency pulse
with a frequency equal to the Larmor frequency is applied perpendicular to the static magnetic
field (B0), causing the protons in nuclei to be stimulated, and spun out of equilibrium. Next, when
the radiofrequency pulse is removed, the nuclei return to the equilibrium and again become parallel
with the static magnetic field. This return is termed relaxation. During relaxation, a radiofrequency
signal is generated by losing the energy of the nuclei. This signal is called the free-induction
decay (FID) response signal. Finally, a conductive field coil positioned around the patient being
imaged is used to measure the FID response signal. This measurement is rebuilt and processed to
acquire three-dimensional grey-scale MRI images [4] [5].
MRI is a conversant imaging technique commonly employed for cancer diagnosis and detection.
Different MRI sequences can be utilized to provide diverse types of imaging information. The
most commonly used MRI sequences for cancer detection are T1-weighted (T1W) sequence, T2-
weighted (T2W) sequence, fluid attenuation inversion recovery (FLAIR), diffusion-weighted
imaging (DWI), and contrast-enhanced T1 (T1CE), [6]. First, T1W is one of the basic pulse
sequences in MRI, also termed spin-lattice relaxation time, and depends on the T1 relaxation time
of tissues. T1 represents the time required for the proton's spins to back to equilibrium, and be
parallel to the static magnetic field. T1W tends to have a short echo time and short repetition time.
Consequently, fats return faster to the longitudinal axis (B0) and appear bright on T1W images.
On the other hand, water takes a longer time for longitudinal magnetization realignment after
applying a radiofrequency pulse and appears dark [7]. Second, T2W, also called spin-spin
relaxation time, depends on the T2 relaxation time (transverse relaxation time), which represents
the amount of time required for the proton's spins to dephase among the nuclei spinning
perpendicular to the static magnetic field. T2W tends to have a long echo time and a long repetition
time. Consequently, water appear bright in T2W images, while fats seem dark [8]. Third, FLAIR
has a long inversion time and is one type of inversion recovery sequence. This sequence eliminates
the signal from the cerebrospinal fluid in the generated images. This can be achieved by adjusting
the inversion time of the FLAIR pulse so that the fluid exhibits no net transverse magnetization at
equilibrium. On FLAIR images, brain tissues seem comparable to T2W images, but Cerebrospinal
fluid is dark instead of bright [9] Lastly, DWI sequence is one type of MRI imaging that depends
on quantifying the water molecules' random Brownian motion within a tissue voxel. Generally,
highly cellular tissues show lower diffusion coefficients. The more individual water molecules
diffused during the sequence, the further it will be revealed to variable gradient strength, and the
more it will be dephased, which will reduce the amount of signal produced. The strength of this
effect is specified by the value of b (a factor that reflects the strength and timing of the gradients
used to generate DWI). Diffusion is mainly employed in cerebral ischemia and tumor
characterization [10]. Figure 1 illustrates the isolation of lesions of the splenium of the corpus
callosum for the aforementioned MRI sequences [11].
Figure 1: Isolated lesion of the splenium of the corpus callosum in a) T1W, b) T2W, c) FLAIR, d) T1 contrast-
enhanced, and e) DWI.
The use of contrast agents (CAs) in MRI images allows the detection of tumors and cancerous
tissues at an early age. Nowadays, researchers are putting efforts into improving the abilities of
MRI in oncology by adding specific contrast agents to cell surface receptors or extracellular matrix
parts, therefore changing the proton relaxation times [12]. These contrast agents lower T1 and T2
relaxation times, resulting in high-intensity signals in T1W images or lower-intensity signals in
T2W images. MRI contrast agents are classified based on the subsequent particular characteristics:
chemical composition, magnetic properties, biodistribution, and imaging application. These agents
can be injected intravenously or delivered orally [13].
One class of the CAs in MRI for cancer detection is the paramagnetic metal Gd3+, which modifies
the static magnetic field partaken by surrounding protons, leading to speeding their relaxation time,
hence yielding high-intensity signals in T1W images. Another class of CAs in MRI is the
superparamagnetic iron oxides (SPIO), which modifies the contrast of MRI images based on
shortening T2 relaxation time. This class is more useful than Gd3+ since it can be used in lower
concentrations. Moreover, the dynamic contrast-enhanced (DCE) agent is one of the CAs in MRI
and is used to detect the changes in perfusion resulting from the establishment of new vessels once
the tumor has grown larger than 2-3 mm. This type of contrast agent is particularly useful in the
detection of prostate and breast adenocarcinomas [14]. Even though the DCE technique has shown
effective detection of small tumors, the sensitivity of this method for detecting malignancy at the
earliest stages is weak. Therefore, many researchers are trying to develop CAs in MRI that are able
to catch slight changes in the intra and extracellular milieu during oncogenesis. This can be
achieved by using targeted CAs MR, which are attached to a targeting moiety, such as antibodies,
peptides, or small molecules, that exist in the extracellular matrix, neovascularization, death cell,
or organ surfaces. For example, a study shows that the use of Gd-chelates contained
in perfluorocarbon nanoparticles as targeted CAs by attaching to α5β3 antibodies was able
to image neovascularization in rabbit cornea and increase the signal intensity by 25% [14].
3. Preprocessing Techniques
In the diagnosis and analysis of cancerous tissues, it is crucial to acquire precise and accurate
images to prevent wrong observations. Thus, the acquired MRI images often pass through several
preprocessing techniques to reduce the noise and to enhance the contrast and the image quality.
There are several common used techniques used to denoise and improve the MRI images quality,
such as Gaussian Smoothing, Bilateral Filtering, and Non-local Means Denoising. Beside the
common used preprocessing techniques, there are several advanced ones being explored for noise
reduction and image enhancement in MRI analysis. In this paper, two advanced methods will be
introduced: Dictionary Learning-based Denoising and Wavelet-based Denoising.
Firstly, Dictionary Learning-based Denoising in MRI is a technique that uses dictionary learning
and sparse coding to decrease noise in MRI images. This method aims to improve the quality and
clarity of MRI images by removing unwanted noise artifacts.
The denoising process involves representing each image patch as a linear combination of atoms
from the learned dictionary. This sparse coding representation allows for the separation of the
signal (desired information) from the noise components. By promoting sparsity in the
representation, the denoising algorithm can effectively eliminate noise while remaining the
important features of the MRI data [15].
The dictionary learning-based denoising approach typically consists of the following steps :
1- Dictionary Learning: This process starts by acquiring a dictionary by learning from a set
of training data that shares similarities with the images intended for denoising. The
dictionary contains a group of basis functions referred to as atoms, which are capable of
representing various image patches. Then, dictionary learning algorithms, such as K-SVD
or the method of optimal directions (MOD), are employed to iteratively refine and improve
the dictionary so that it optimally matches the training data.
2- Sparse Coding: Each image patch is represented as a sparse linear combination of atoms
from the learned dictionary. The sparse coding step promotes sparsity in the representation,
allowing for effective noise suppression .
3- Denoising: The sparse coding representation is used to separate the signal (desired
information) from the noise components. By removing the noise components, the denoised
MRI data is obtained .
Secondly, Wavelet-based denoising algorithms use the mathematical concept of wavelet transform,
which decomposes the image into several frequency bands. This algorithm exploit the ability of
wavelets to represent both smooth and detailed features of the image. By analyzing the image in
different frequency scales, this approach can effectively separate noise from the underlying image
structure.
When an image denoised using wavelet transformation, it results in four sub-images , as shown in
Figure 2, known as approximation details, horizontal details, vertical details, and diagonal details.
These sub-images have two types of coefficients: small coefficients and large coefficients.
Figure 2: Wavelet transform components (a) Horizontal (b) Vertical (c) Diagonal [16].
The small coefficients are primarily caused by noise and can be effectively eliminated without
affecting the important features of the image. The large coefficients ,on the other hand, represent
essential signal features.
In other techniques like thresholding or non-linear techniques, only one wavelet coefficient is
processed at a time. In thresholding, each coefficient is compared to a predetermined threshold
value. If a coefficient is smaller than the threshold, it is set to zero; otherwise, it is retained or
modified.
After the thresholding process, an inverse wavelet transform is performed to reconstruct the image
with reduced noise while preserving the important characteristics of the image [16].
4. Segmentation of Cancerous Tissues
It's crucial to perform image segmentation, where regions of interest are outlined, for complete
visualization and investigation. Segmentation is mainly the procedure of dividing an image into
pieces having a strong correlation at regions of interest in this image. It aims to modify or simplify
the overall presentation of the image to become more expressive and easily analyzed. Different
techniques can be employed for image segmentation. However, not all approaches are suited for
medical purposes, as they are complex and inaccurate. Otsu’s, watershed, K-means, HAAR
Discrete Wavelet Transform (DWT), and Convolutional Neural Network (CNN) are examples of
image segmentation algorithms in MRI, particularly for cancerous tissue detection [17].
First, Otsu’s is a nonlinear image thresholding-based segmentation technique, which is employed
to convert an image from grayscale into a binary. This approach generates a prime threshold by
minimizing all weighted sums of one-class variances of foreground and background pixels. Then,
the statistical technique of stationary objects is utilized to pinpoint the items, rather than spatial
coherence [17]. Second, the watershed technique utilizes topographic information for dividing the
image into segments. It deals with the image as a topographic surface by specifying catchment
basins based on pixel intensity. Local minima are characterized as beginning points. Then, flooding
with colors refills catchment basins until object borders are obtained [18]. Third, the K-mean
clustering algorithm is an unsupervised approach that categorizes the input data points into
multiple categories based on their intrinsic distance from each other, and it gives prime
segmentation for the image [17]. Finally, CNN is commonly used in the segmentation of brain
tumor images, in which the information moves from the right side of the model without any
feedback. CNN is processed through three different layers: a sequential convolutional layer,
pooling layers, and connected layer. The convolutional layer is utilized to map the characteristics
of the image with the convolution kernel. The pooling layer is used for subsampling by applying
the pooling function to a spatial window. Next, the output obtained from the pooling layer
is then coupled with the connected layer, giving the feature vector [17].
According to Adesh, the findings of simulating the aforementioned segmentation techniques in
MATLAB software revealed that the estimated accuracy of Otsu’s, watershed, K-means, and CNN
methods is 71.42%, 78.26%, 84.34%, and 91.39 respectively. Moreover, the response time of CNN
was 2.519 s, indicating that the CNN segmentation algorithm for cancerous tissue detection is the
best in comparison to all other approaches [17].
5. Machine learning for cancer detection using MRI.
Machine learning has become essential in the process of cancer detection in MRI as it can perceive
cancer symptoms and notice them where people cannot. It is also important for the large amount
of data as it offers a fast and accurate analysis system compared to manual analysis. Machine
learning algorithms can produce models that classify MRI images into malignant and benign
according to local and global image features. Those models are considered mathematical models
that can apply their stored set of data to estimate a result. The learning exercises can be either
supervised, semi- supervised or unsupervised. However, imaging data is usually used in supervised
learning exercises. The following table describes three major machine learning models diagnostic
characteristic in cancer detection using MRI.
Table 1: ML models Vs Diagnostic characteristics [18]
Model Diagnostic characteristics
Artificial neural network (ANN) • Neural networks (NNs) use non-
linearities and weights to handle
complex problems by transforming
and combining input data.
• Training a NN requires multiple
iterations over the training data, which
can take a long time.
• Tuning hyperparameters, such as the
number of hidden layers and nodes, is
necessary to achieve optimal
performance in NNs
Decision trees (DTs) • Easy to visualize and comprehend.
• One set of features can provide
entirely different performance than
another set of features, so a large
Random Forests can be used to avoid
this problem.
• Prone to overfitting.
Contrastive learning • Self-supervised
• task-independent deep learning
technique
• Learns the general features of a
dataset by differentiating similar and
different images
6. Clinical Application and Challenges
MRI is widely used in cancer conditions to detect and diagnose the disease. For example, the tumor
needs a blood supply to grow and spread, so a new blood vessel will be formed to provide the
needed blood and this process is called Angiogenesis. In this case, MRI could detect the blood
supply that is concentrated to the tumor. Also, MRI could provide accurate information about the
tumor characteristic and stage such as its place, spread, size, shape, and extent.
It is clear that MRI is important and valuable in cancer detection, but also there are many
challenges that must be taken into consideration. First, MRI is known to be expensive with a high
equipment and maintenance cost, and it takes a long time to be performed. In addition, MRI cannot
be used for all patients such as patients with implanted pacemakers or other medical devices that
contain magnetized metal parts. Also, Anxiety considered one of the common challenges in MRI
that cause the patient to move, which could damage the picture or harm the patient [19].
7. Conclusion
To conclude, MRI-based cancer detection offers great promise in providing early diagnosis and
treatment of cancer. Using a balance between the power of evolved imaging techniques, such as
feature extraction and machine learning algorithms, it can increase accuracy and efficiency in
recognizing cancerous tumors. However, challenges such as data variability, limited availability of
labeled datasets, and interpretability of deep learning models constantly arise and need to be
solved. With further research and technological advancements, MRI-based cancer detection has
the potential to revolutionize cancer diagnosis and improve overall healthcare.
8. References
[1] “What is cancer?: Cancer basics,” What is Cancer? | Cancer Basics | American Cancer
Society, [Link] (accessed
Apr. 28, 2024).
[2] “Cancer,” Mayo Clinic, [Link]
treatment/drc-20370594 (accessed Apr. 28, 2024).
[3] “MRI for cancer: Magnetic Resonance Imaging test,” Magnetic Resonance Imaging Test |
American Cancer Society, [Link]
tests/[Link] (accessed Apr. 28, 2024).
[4] Magnetic resonance Imaging (MRI). (n.d.). National Institute of Biomedical Imaging and
Bioengineering. [Link]
imaging-mri
[5] 2.2 Basic principles of MRI. (n.d.). [Link]
[6] Kronberg, R. M., Meškelevičius, D., Sabel, M., Kollmann, M., Rubbert, C., & Fischer, I. (2022).
Optimal acquisition sequence for AI-assisted brain tumor segmentation under the constraint of largest
information gain per additional MRI sequence. Neuroscience Informatics, 2(4), 100053.
[Link]
[7] Baba, Y., & Jones, J. (2009). T1 weighted image. [Link]. [Link]
[8] Haouimi, A., & Jones, J. (2009). T2 weighted image. [Link]. [Link]
6345
[9] Baba, Y., & Niknejad, M. (2013). Fluid attenuated inversion recovery. [Link].
[Link]
[10] Goel, A., & Bashir, U. (2012). Diffusion-weighted imaging. [Link].
[Link]
[11] JNC :: Journal of Neurocritical Care. (n.d.). [Link]
[Link]&id=f1-jnc-9-1-21&number=221&p_name=0717_221
[12] JNC :: Journal of Neurocritical Care. (n.d.). [Link]
[Link]&id=f1-jnc-9-1-21&number=221&p_name=0717_221
[13] Xiao, Y., Paudel, R. C., Li, J., Ma, C., Zhang, Z., & Zhou, S. (2016). MRI contrast agents:
Classification and application (Review). International Journal of Molecular Medicine, 38(5), 1319–1326.
[Link]
[14] GeeksforGeeks, “Image denoising using dictionary learning in Scikit Learn,”
GeeksforGeeks, [Link]
learning-in-scikit-learn/ (accessed May 4, 2024).
[15] “[PDF] analysis and comparison of wavelet transforms for denoising MRI image,” Semantic
Scholar,
[Link]
(accessed May 4, 2024).
[16] Bogdanov, A. A., & Mazzanti, M. L. (2011). Molecular magnetic resonance contrast agents for the
detection of cancer: past and present. Seminars in Oncology, 38(1), 42–54.
[Link]
[17] Ullah, F., Nadeem, M., Abrar, M., Al-Razgan, M., Alfakih, T., Amin, F., & Salam, A.
(2023). Brain Tumor Segmentation from MRI Images Using Handcrafted Convolutional Neural
Network. Diagnostics, 13(16), 2650. [Link]
[18] GeeksforGeeks. (2024, February 12). Image Segmentation with Watershed Algorithm OpenCV
Python. GeeksforGeeks. [Link]
opencv-python/
[19] David, “Common challenges faced by MRI technologists & how to overcome them,”
MRIaudio, [Link]
overcome-them/ (accessed Apr. 29, 2024).
