1.

Breakthrough in AI for Breast Cancer Detection: Bridging

Privacy and Power
1. Author Names:
• Hussain AlSalman
• Mabrook S. Al Rakhami (main contact)
• Taha Alfakih
• Mohammad Mehedi Hassan
2. Objectives
• Beating the Small Data Problem:
AI models that are trained using information from one hospital are not very good. The main idea was
to make a model that can learn from a lot of hospitals at once. This way the AI model will be better
because it learns from hospitals combined. The goal is to have a model that learns from hospitals so it
will be stronger, than AI models that only learn from one hospital.
• Protecting Patient Privacy:
People wanted hospitals to work together without sharing sensitive medical images. The main goal
was for hospitals to collaborate on images without actually sharing the medical images.
• Introducing a New, Robust Solution:
The people, behind this project proposed a way of doing Federated Learning, which is called
Federated Learning and it is specially made for medical imaging. They want to use this Federated
Learning method to help with medical imaging.
3. Existing methods
The usual way to find breast cancer is by using special computer programs called Deep Convolutional
Neural Networks. These Deep Convolutional Neural Networks are very good at looking at pictures
and finding breast cancer. They are, like smart machines that help doctors see things that are not easy
to see. The doctors use these Deep Convolutional Neural Networks to look at the pictures and find the
breast cancer.
• The Method: Hospitals and researchers will use these Deep Learning Networks to look at pictures
like the ones women get when they have a mammogram to see if they can find cancer. The Deep
Learning Networks get smarter by looking at a lot of pictures of images and that is how the Deep
Learning Networks learn.
• Deep learning computers are very accurate. Help doctors make decisions faster and get it right more
often. Doctors use these computers to look at things like cancer. They are a big help. The computers
are good, at finding cancer.
4. Gaps:
• more Data
Artificial intelligence needs a lot of pictures actually millions of them to be really good, at what it
does. The more images artificial intelligence has the better artificial intelligence can do its job.
Artificial intelligence uses all these images to learn and become highly accurate.
• Patient Privacy
Laws like HIPAA prevent hospitals from sharing patient images.
Hospitals cannot share information with each other. This means that each hospital has to train its
artificial intelligence model using only the pictures it has. The artificial intelligence model that each
hospital trains is not very good because it does not have pictures to learn from.
The big question is how to create an artificial intelligence model that all hospitals can use without
sharing private information about patients.
5. Proposed Methods
• Share insights:
Hospitals never share data with anyone. They only share the insights that they get from the
information. The hospitals use their local models to learn things and they only share what they have
learned from patient data. They do not share the patient data itself. Hospitals keep data to themselves
and only share the insights their local models learn from this data.
•.The computer only sends updates from the model to a server. It does not send the pictures
Two Key Components:
• Dual Head Deep Convolutional Neural Network (DH DCNN):
A custom AI model designed to analyze breast cancer images more effectively.
• Dynamic Aggregation Scheme (DAS):
A smart method used by the central server to combine updates from all hospitals and build the
strongest global model.

________________________________________

2. Deep Learning and Federated Learning in Breast Cancer

Screening and Diagnosis: A Systematic Review
[Link] Names
• Alexandru Ciobotaru
• Cosmina Corches
• Dan Gota
• Liviu Miclea (The main contact for the research)
2. Objectives
• Provide a Clear Roadmap: want an detailed plan that explains all the new developments in breast
cancer screening and diagnosis. This summary should have all the information on how Deep Learning
and Federated Learning are being used to improve breast cancer screening and diagnosis.
3. Existing Methods
The paper reviews two existing technological approaches to breast cancer diagnosis that have already
shown significant progress in reliable and secure breast cancer diagnosis.
• Deep Learning (DL): The use of advanced neural networks for analyzing medical images.
• Federated Learning (FL): The decentralized method of training AI models across multiple data
sources while keeping data local.
4. Gap
• We Are Missing The Big Picture: The field of research is changing fast so people who study it need
a simple and complete summary of all the new discoveries and what is happening now with the
current trends, in the field.
• Uncertainty in Direction: Researchers are having a time figuring out what to do next because there is
no clear summary of what has already been done. This makes it really hard for them to know the way
to move forward with their work. So a roadmap of research directions for the researchers was really
needed to help them out.
5. Proposed Methods
• The Method: The idea here is to do a Systematic Review. This means we will collect all the studies
that're relevant look at them carefully and then put everything together to get a complete picture of
what is going on in the field of Systematic Review.
________________________________________

3. A Transfer Learning Approach to Breast Cancer Classification

in a Federated Learning Framework
1. Author Names
• Y. Nguyen Tan
• Vo Phuc Tinh
• Pham Duc Lam
• Nguyen Hoang Nam
• Tran Anh Khoa
2. Objectives
The team wanted to make traditional AI training better. They also wanted to improve the power of the
AI training. The team aimed to:
• We need to stop putting all our information in one spot because that is a big risk for our privacy.
When we gather all the information in one place to train models it is very likely that someone will get
unauthorized access to it and that will be a huge problem, for patient privacy. Patient data is very
sensitive. We need to protect it from people who want to misuse it.
• We need to make a computer program that can find and locate bad cells in the body and tell us if
they will become tumors. This will help doctors take care of patients with cancer and increase the
chances of Patient Survival. The goal is to improve Patient Survival by using this technology to make
more accurate decisions, about treatment.
3. Existing Methods
• Centralized Learning: When it comes to breast cancer diagnosis most artificial intelligence
applications have been made in a Centralized Learning setting. This is where all the information is
gathered and kept in one place, which's the Centralized Learning environment. The thing about
Centralized Learning is that it stores all the data, in one spot.
4. Gap
• The Privacy Time Bomb: The thing is, most artificial intelligence research uses Centralized
Learning, which's the Privacy Time Bomb and this Centralized Learning puts sensitive patient
information at risk of being seen by people it should not be seen by which is a big problem, for patient
privacy.
• The Need for Precision: We really need to make sure that artificial intelligence is good at finding and
pointing out the location of abnormal areas, within the breast tissue so that we can do a better job of
predicting tumors in the breast tissue.
5. Proposed Methods
• The Method: The researchers came up with a way to use something called Transfer Learning for
breast cancer classification. They did this by using Transfer Learning, inside another system called
Federated Learning.. They used Transfer Learning and Federated Learning together to try to improve
breast cancer classification. This way of doing things lets the Artificial Intelligence use what it already
knows, which is called Transfer Learning. It also gets to learn from lots of different sources of
information called Federated Learning in a really safe way.
________________________________________

3. Enhancing Breast Cancer Classification in Histopathological

Images through Federated Learning Framework
1. Author Names
• Jyothi Peta
• Srinivas Koppu (The main contact for the research)
2. Objectives
The main goal was to make a tool to help doctors find out if someone has breast cancer. They wanted
to do a things like:
• We should make a system that can automatically diagnose diseases. This system will use something
called learning and deep learning to make the diagnosis. The idea is to use learning and deep learning
to create an automated disease diagnosis system that can really help people. We will use learning and
federated learning to make this automated disease diagnosis system work.
• Increase Efficiency: Automate and speed up the process of detection efficiently.
• Make A Good System: We need to make a strong system to find breast cancers from the samples we
get. This system has to be really good at finding breast cancers.
3. Existing Methods
• There were some studies before that tried to find a good way to detect breast cancers from tissue
samples. These studies on breast cancers were done to see if we can find breast cancers early. The
people doing these studies, on breast cancers wanted to make sure they can detect breast cancers from
the tissue samples of breast cancers.
4. Gap
• Low Effectiveness: The things that were tried before did not work well. These prior attempts had
limitations, which made the Low Effectiveness of these attempts very clear. The Low Effectiveness
was a problem because it meant that the prior attempts were not good enough.
• The Security Crisis: When we talk about the sector one big problem is that medical images are very
sensitive and it is hard to share them in a safe way. The medical sector is still trying to find a solution,
to this problem of sharing these medical images.
5. Proposed Methods
• The Method: This is a system that uses computers to help doctors diagnose diseases. It combines
two things: a way for computers to work together called Federated Learning and a smart way for
computers to learn called Deep Learning. The system is called an automated disease diagnosis system
because it uses these two things to help figure out what is wrong with someone. The automated
disease diagnosis system is really good, at helping doctors diagnose diseases.
• The system they have, in place includes five steps that are meant to keep everything safe and
working properly:
1. Image Acquisition
2. Encryption
3. Optimal Key Generation
4. The Homomorphic Encryption application is a kind of encryption. This Homomorphic Encryption
lets people work with data that is already encrypted. So with Encryption we can do things with the
encrypted data without having to decrypt it first. This is what makes Homomorphic Encryption so
useful. The Homomorphic Encryption application is really good, for keeping data safe.
5. Classification
________________________________________

4. Breast Cancer Prediction Using Shapely and Game Theory in

Federated Learning Environment
1. Author Names
• Y. Supriya
• Rajeswari Chengoden (The main contact for the research)
2. Objectives
The team worked on making the privacy of the AI model also on making the internal mechanics of the
AI model better. The team had some goals, for the AI model these were to:
• Propose a framework: We need to create a system that brings together Shapley values and game
theory with Federated Learning. This system should be smart. Work well with Federated Learning.
We will call this system a combination of Shapley values and Federated Learning and it will help us
make better decisions.
• We want to find breast cancer. This will help us get results when we try to predict if someone has
breast cancer. If we can do that we can save lives and help people with breast cancer live longer. We
need to improve our methods, for detecting breast cancer so we can help people with breast cancer.
• Make Sure Peoples Information Is Safe: We need to find a way to teach computers to make
predictions using health information from different places without putting peoples private information
at risk. We have to make sure that the health information of people is kept private and secure when we
use it to train computers to make predictions about health. This is very important, for health care data.
3. Existing Methods
• Traditional AI: The usual methods of Machine Learning and Deep Learning have worked well for
predicting breast cancer.
4. Gap
• Federated Learning is what we need to solve the problem of keeping healthcare data secure. The
main issue is that we need to make sure that healthcare data is used to train these models without
putting the data at risk. This is where Federated Learning comes in to help with the Privacy and
Performance issue.
5. Proposed Methods
• The Method: We have a way of doing things that combines ideas from Shapley values and Game
Theory with Federated Learning.
• This is how it works:
Shapley Values: Used as a tool for feature selection to pick the most important variables from patient
data (specifically, 30 features from the Wisconsin Diagnostic Breast Cancer dataset).
Game Theory: Integrated with FL to improve the collaborative learning process.

5. A Novel Federated Learning Framework for Sustainable and

Efficient Breast Cancer Classification System (FL-L2CNN-
BCDet)
1. Author Names
 NESMA ABD EL-MAWLA (The main contact for the research)
 MOHAMED A. BERBAR
 NAWAL A. EL-FISHAWY
 MOHAMED A. EL-RASHIDY
2. Objectives
The main goals of this project were to make the intelligence that helps detect breast cancer better. We
wanted to make the artificial intelligence used for breast cancer detection more responsible. This
means we focused on making breast cancer detection using intelligence more sustainable and
efficient.
 Improve Data Security and Privacy: Secure patient data from breaches.
 Improve The Cancer Detection Rate: We need to make the Deep Learning model a lot
better at finding cancer. The Deep Learning model has to be more accurate when it is looking
for cancer.
 Minimize Waste: We need to create a system that uses fewer resources. This means we have
to reduce the amount of energy consumption and make the system latency better so it does
not have a slow response time. We have to focus on reducing energy consumption and
latency to make this work. By doing this we can make the system more efficient.
3. Existing Methods
 Centralized Learning (CL): A lot of the research on Artificial Intelligence uses Centralized
Learning. This means that Centralized Learning is the method where all the patient data is
gathered together and kept in one spot, for training purposes. Centralized Learning involves
collecting all the data in one place.
 The Problem with Centralized Learning (CL): The thing with this way of doing things,
where everything is controlled from one place, is that it is really bad for keeping patient
information private. Centralized Learning is not good, at protecting patient information.
4. Gaps (The Hurdle)
 Privacy Risks: The thing is, when we use Artificial Intelligence solutions they put patient
privacy at risk because they store all the data in one place and that is the main problem, with
Artificial Intelligence solutions.
 Need for Efficiency: The researchers needed a system that could give them very good results.
This new system also had to be sustainable. This means it had to use less power and it had to
be faster. The researchers wanted a system, without high latency. They wanted the system to
work well and not use too much power. The sustainable system had to be efficient.
5. Proposed Methods
 The Method: The people who made this Federated Learning framework for Breast Cancer
classification came up with an idea. They called it FL-L2CNN-BCDet.
 Key Component: The framework uses a kind of artificial intelligence model called FL-
L2CNN. At the same time it helps save power. The people who made FL-L2CNN wanted to
make sure it was very accurate. They also wanted it to use less energy.
6. Lung-AttNet: An Attention Mechanism-Based CNN
Architecture for Lung Cancer Detection With Federated
Learning
1. Author Names
 CHAMAK SAHA
 SOMAK SAHA
 MD. ASADUR RAHMAN
 MD. MAHMUDUL HAQUE MILU
 HIROKI HIGASHI
 MOHD ABDUR RASHID
 NASIM AHMED (The main contact for the research)
2. Objectives
The team wanted to make lung cancer detection better. They wanted it to be more reliable and private.
The team also wanted it to be able to handle image data from lung cancer scans.
 Protect Patient Data: Address the serious challenges related to privacy and security when
sharing sensitive medical data across institutions.
 Boost Detection Accuracy: This model has to be very good at finding the cancer so we can
help people with lung cancer. We are talking about using these pictures to make the model
better at finding lung cancer.
 Overcome Data Shortages: We need to solve the issue where each hospital only has a
limited amount of good patient data that is labeled correctly. This is a problem because
hospitals like these hospitals need a lot of patient data to do their jobs well. The patient data
that these hospitals have is just not enough. We have to find a way to get more patient data for
these hospitals so they can work better and help people. Overcome Data Shortages is very
important for these hospitals and, for the people who go to them.
3. Existing Methods
 Traditional Deep Learning: Traditional Deep Learning models need people to share their
data with a central place.
 The Problem with Centralized Data: When we centralize data it is a problem because it
goes against the rules that are supposed to protect our privacy. This also means that the
artificial intelligence models we create from this data are not very good at working in other
hospitals like the hospital, down the street because they do not work well with different types
of data, which is known as poor generalization of the centralized data and the artificial
intelligence models.
 Older Detection Tools: The old ways of finding lung nodules are not very good. They often
say something is wrong when it is not. They take a long time to give an answer. This happens
because these older lung nodule detection tools are not perfect. These older lung nodule
detection tools have a lot of limitations.
4. Gaps
 Privacy and Generalization: The main problem is that we need a lot of information to make
Artificial Intelligence work well but we also need to keep peoples information private and
safe. If we only use limited information from one place to train Artificial Intelligence it will
not work reliably everywhere with Artificial Intelligence.
 Accuracy Issues: The tools we have now to find lung nodules take a long time and
sometimes they tell people they might have cancer when they really do not have cancer. This
happens a lot with lung nodule detection tools, they make a lot of mistakes. 5. Proposed
Methods
 The Method: A novel privacy-preserving framework combining Federated Learning (FL)
with a new model architecture called Lung-AttNet.
 Key Component: The Lung-AttNet is a kind of computer program called an Attention
Mechanism-Based Convolutional Neural Network or Lung-AttNet for short. This Lung-
AttNet attention mechanism is like a tool that zooms in on the parts of a picture, like a tiny
dot that could be cancer. The Lung-AttNet helps the computer focus on these dots, which are
the parts of the image that really matter.

7. Joint Federated Learning Using Deep Segmentation and the

Gaussian Mixture Model for Breast Cancer Tumors
1. Author Names
 NGUYEN TAN Y
 PHAM DUC LAM
 VO PHUC TINH
 DUY-DONG LE
 NGUYEN HOANG NAM
 TRAN ANH KHOA (The main contact for the research)
2. Objectives
The main idea was to make the Artificial Intelligence better, at finding the shape of cancer tumors.
This had to be done in a way that kept the process safe and secure for the people involved with the
cancer tumors.
 Improve Segmentation Accuracy: We need to get really good at image segmentation. This is
the process of drawing an outline, around the tumor. Medical image segmentation is very
important for doctors to make clinical decisions. They rely on medical image segmentation to
get it right. So we have to improve image segmentation accuracy.
 Address Data Challenges: Resolve the complex challenges caused by two main factors:
having images from very different sources and having major data sharing and privacy
concerns.
  Develop a Secure System: We need to create a system that's really secure. To do this we
should work together on a kind of learning approach that combines the best methods to make
sure the system is both very accurate and completely safe, for our data. 3. Existing Methods
 Deep Learning Segmentation: We use Deep Learning models to find the edges of tumors in
medical pictures. This is called deep learning segmentation. Deep learning segmentation is
really important because it helps doctors see the tumors clearly in these pictures
4. Gaps
 Data Heterogeneity (Diverse Sources): Medical images come from lots of hospitals and
they use different machines to take these pictures. This makes it really tough to create one
model that can correctly identify what is in all these images, from Data Heterogeneity. The
Data Heterogeneity is a problem because the pictures are all different.
 Data Sharing and Privacy: When we talk about Data Sharing and Privacy we have to think
about how Data Sharing and Privacy affects us. The big problem with Data Sharing and
Privacy is that we have to share information so that we can train everything in one place. This
is a hurdle when it comes to making things more accurate, while protecting privacy.
5. Proposed Methods
 The Method: A Joint Federated Learning (FL) approach.
 Key Components: This joint approach brings together two key components and advanced
techniques:
1. Deep Segmentation: This is a tool that helps doctors find the shape of a tumor. It is
really good at drawing a line around the tumor. The Deep Segmentation model is
very useful, for professionals who need to know the exact size and shape of a tumor.
2. Gaussian Mixture Model (GMM): Used to classify and interpret the results of the
tumor segmentation.

8. Bilevel Hyperparameter Optimization and Neural

Architecture Search for Enhanced Breast Cancer Detection in
Smart Hospitals Interconnected With Decentralized Federated
Learning Environment
1. Author Names
 SALABAT KHAN (Co-corresponding author)
 FARIHA NOSHEEN
 SYED SHEHRYAR ALI NAQVI
 HARUN JAMIL
 MUHAMMAD FASEEH
 MURAD ALI KHAN
 DO-HYEUN KIM (The main contact for the research)
2. Objectives
The main goal was to make the Federated Learning process really smart. This was done by
automating the design of the Artificial Intelligence model itself. The Federated Learning process was
made smart by doing this.
 Build the Best AI Model: Develop the absolute best possible AI structure (Convolutional
Neural Network) for breast cancer detection.
 Boost Performance in FL: Improve the efficiency and speed of the learning process within
the decentralized Federated Learning (FL) setting.
 Keep The Hospitals Safe And Private: We need to make sure that the hospitals where
people go to learn and get better are very safe and private. This is especially important
because these hospitals are all connected to each other. The security and privacy of the
learning environment, at these hospitals must be very strong.
3. Existing Methods
 Traditional CNNs and FL: People use Convolutional Neural Networks (CNNs) in the
existing Federated Learning (FL) systems. They use these CNNs in traditional FL
frameworks.
 The Problem: Traditional Federated Learning has a hard time when the information from
different hospitals is really different. The data, from these hospitals is just not the same.
Finding the way to set up the network and getting all the settings right is usually done by
hand, which is a very slow process and it does not work very well. Federated Learning faces
these issues because the data is heterogeneous.
4. Gaps (The Hurdle)
 Optimization Challenge: Finding the best network architecture, which is the structure of the
Artificial Intelligence model, and the best hyperparameters, which are the Artificial
Intelligence model's settings, is a really tough and time-consuming thing to do. It takes a lot
of work to get it right. The Artificial Intelligence model needs to be perfect. The settings need
to be just so. This is something that people usually have to do by hand which takes a long
time. Finding the Artificial Intelligence model and the best settings, for the Artificial
Intelligence model is a big challenge.
 Heterogeneity in Federated Learning: Federated Learning has a hard time making a really
good global model when the data used at the different hospitals is all very different. This is
because the data is not the same which is called non-IID data. The problem is that Federated
Learning deals with this kind of non-IID data, which is a big challenge, for Federated
Learning.
5. Proposed Methods
 The Method: This is a way of doing things that combines Bilevel Hyperparameter
Optimization (BHO) and Neural Architecture Search (NAS), in a Decentralized
Federated Learning (FL) Environment.
 This Is How It Works:
o Neural Architecture Search (NAS): This tool is really helpful because it
automatically finds the best possible structure for an Artificial Intelligence model. It
does this by itself so we do not have to try a lot of things to see what works. This
 means we can save time and get an Artificial Intelligence model without having to
guess what will work best. The Neural Architecture Search tool is very useful, for
creating Artificial Intelligence models.
o Bilevel Hyperparameter Optimization (BHO): The Bilevel Hyperparameter
Optimization tool is really helpful. It automatically finds the best settings, which are
also called hyperparameters, for the model. This makes the model learn things faster.
It is more efficient. The Bilevel Hyperparameter Optimization tool is very useful, for
the model because it makes the learning process easier.