Browsing by Author "Chakrabarty, Amitabha"

Now showing 1 - 20 of 172

A comparative study of lung cancer prediction using deep learning
(BRAC University, 2022-09) Mugdho, Aka Mohammad; Bhuiyan, Md. Jawad Hossain; Rafin, Tawsif Mustasin; Amit, Adib Muhammad; Chakrabarty, Amitabha; Rasel, Annajiat Alim
At the point when cells in the body develop out of control, this is alluded to as cancerous development. Lung cancer is the term used to depict cancer that starts in the lungs. At first in the field, classifier-based approaches are joined with various division calculations to utilize picture acknowledgment to recognize lung cancer nodules. This study found that CT scan images are more reasonable for delivering improved results than other imaging modalities. The use of the images is a piece of chiefly inspecting the CT scanned images that are viewed as informational collections for patients affected by lung cancer. The suggestion of our paper exclusively centers around the execution of concentrating on the calculation’s accuracy in diagnosing lung cancer. Thus, the primary plan of our examination is to utilize examined calculations to conclude which strategy is the most efficient method for detecting lung cancer initially. After training the model we found that Over all accuracy of Resnet-18 is 99.54%, the Overall accuracy of Vgg-19 is 96.35%, The overall accuracy of MobileNet V2 is 98.17%, Dense Net161 is 99.09% and Inception V3 is 98.17%. So we can see that ResNet18 perform better than other train model.
A comparison of deep learning U‐Net architectures for semantic segmentation on panoramic X-ray images
(BRAC University, 2023-01) Bin Mushfiq, Rahil; Zannah, Rafiatul; Bashar, Mubtasim; Alam, Md. Nafidul; Rahman, MD Aftabur; Chakrabarty, Amitabha
Digital image processing utilizes deep learning to tackle challenging issues such as image colourization, classification, segmentation, and detection. The medical image analysis field is developing day by day, and segmenting organs, diseases, or abnor malities is a challenging task to complete. Dental disease diagnosis is one of these fields where image segmentation can help gain significant improvements as dentists worldwide face various problems in diagnosing dental diseases with the naked eye. Compared to other medical images, dental radiographic images provide multiple challenges in terms of processing, making segmentation a more complex task. Deep neural network models are used more frequently for various image segmentation ap plications. U-Net is one such model. Multiple variations and advancements have been created for this network model to serve better performance, mainly on seman tic segmentation of medical images. However, comparative studies must determine how well these variants perform in segmenting dental x-ray images. This research uses six U-Net architecture (Vanilla U-net, Dense U-net, Attention U-net, SE U net, Residual U-net, R2 U-net) variants for segmenting dental radiographic X-rays that are extensively and effectively compared. Some U-Net architectural variations under consideration still need to be evaluated for segmenting dental radiographic X-rays. For all architectures, we used 2 and 3 convolutional layers. We used four types of matrices to compare the models: Accuracy, Dice coefficient, F1 score and IoU. Among the variants, Vanilla-unet with two convolutional layers provided the best Accuracy of 95.56% and IoU score of 88% on the validation set for much lesser time than other architectures. On the other hand, when we use three convolutional layers, dense-unet provides the best Accuracy of 95.94% and IoU score of 89.07% on the validation set. However, most of the examined architectures throughout the dataset showed minor changes when segmentation performance was measured using all four accuracy metrics. This study indicates that U-Net is enough for radio graphic X-ray segmentation. Choosing simpler models will save time and money during testing and model creation. Therefore, our suggested approach might aid in making automated dental disease diagnosis models.
A comprehensive model for advanced road scene understanding: YOLO-CNN fusion for accurate road segmentation and object detection in varied conditions
(BRAC University, 2025-01) Rahman, Washifur; Roy, Shamir; Islam, Ahsanul; Chakrabarty, Amitabha
The critical component in autonomous driving is road scene understanding that includes accurate object recognition and precise road segmentation. Most existing models do not accurately represent the roads of Bangladesh as those are are mostly trained and tested on the data from the organized roads found in developed countries. This study tackles the challenging problem of road perception in the context of autonomous driving in Bangladesh by proposing a YOLO-CNN fusion model that takes into account the country’s diverse weather patterns and unstructured road layouts. To evaluate model performance, we tried out various object detection architectures, such as YOLOv8, YOLOv9, YOLOv10, and YOLOv11, and for road segmentation, we tested U-Net and ResU-Net segmentation models as well. We gathered a dataset of 2,495 images of roads from different locations and environments including highways, villages, foggy and rainy situations, as well as nighttime. The results of the experiments confirmed that the Medium model of YOLOv11 was the most accurate of all versions of YOLO at object detection with an accuracy of 79.3%. Likewise, U-Net outperformed ResU-Net in terms of accuracy and the IoU score of 80.69% for road areas, indicating a closer match to actual road areas. The best-performing models of object detection and segmentation are then combined to create a comprehensive road scene understanding system. The findings show that the combination of YOLOv11 and U-Net Fusion Model enhances object detection segmentation in road environments greatly, which makes it relevant for self-driving car applications. The system was additionally implemented as a web based prototype so that users could upload images and see the results of detection and segmentation visually. Work on these objectives will emphasize improving the economization of processing power for use with low-resource devices, including depth perception with LiDAR, and widening the dataset for better performance across different driving environments.
A deep learning approach to integrate human-level understanding in a Chatbot
(2021-01) Al Mamun, Amirul Islam; Abedin, Afia Fairoose; Nowrin, Rownak Jahn; Chakrabarty, Amitabha
AI-powered computers like chatbots have taken over the market today to reduce human workload. Unlike humans, chatbots reply immediately, are available 24/7 and can assist several people at the same time. Due to the outbreak of Covid-19[3] as everything has just shifted to online, the demand of bots has increased tremendously. Considering the various applications, it is estimated that the chatbot market will reach $1.25 billion by 2025[2]. Though chatbots perform well in task-oriented activities, in most cases they fail to understand personalised opinions, statements or even queries. Generally, people prefer human agents as it is easier to share personal views and feedbacks with them. Also, poor understanding capabilities of a machine disinterest humans to continue conversations with them. Usually, chatbots give absurd responses when they are unable to interpret a user’s text accurately. Hence, it is very essential to develop chatbots with human-level understanding. Most bots are incorporated with sentiments to analyse reviews of products and services of an organisation. However, this is not enough as only positive and negative judgements cannot help an organization improve their lackings. To make chatbots function more precisely, it needs to identify the granular reaction of a customer as well as the reason behind it. Thus, in our research we incorporated all these key features that are necessary for a chatbot to have a human-like understanding of a text. We performed sentiment analysis, emotion detection, intent classification and name-entity recognition using deep learning to modify chatbots with humanistic understanding and intelligence. Conventionally, machine learning is used to perform analysis of the components mentioned above, however, it is seen that ML models ,often ,are unable to understand the inferences and complex sentences of human utterance and this is where deep learning has the upperhand [1]. Therefore, we chose deep learning models such as LSTM, Bi-directional LSTM, GRU, Bi-directional GRU etc to train our chatbot so that it can make more accurate predictions. From our training, we got the best performance model LSTM with accuracy 89% in sentiment analysis and Bi-directional GRU with accuracy 91%, 80.7%, 98.9% for emotion detection, intent classification and named-entity recognition respectively.
A deep learning approach towards soft biometrics attributes prediction using CNN
(BRAC University, 2021-09) Kibria, Maharab; Tabassum, Ilmi; Ahmed, Fardin; Habib, Nahian; Chakrabarty, Amitabha
Any physical, behavioural or adhered human characteristics that we can observe from a person is known as Soft Biometric.The most common physical soft biometric attributes are height, age, ethnicity, facial hairs, gender, hair color etc. In this era of machine and deep learning, retrieving a person based on these semantic descriptions has become a major research interest. Face recognition and bounding boxes are now common implementations in IoT and surveillance systems because of the efficiency of training models. But the research on soft biometric attributes training models still lacks an amount. To overcome this, we have trained different CNN models for the best outcoming prediction result with a UTKface dataset. The dataset includes height, age and gender and 48x48 text pixels face images. The models include CNN, Multi-Headed CNN, DenseNet-169, Multi Label CNN and ResNet- 50. After training all the models we have found that the DenseNet-169 model can achieve the most accuracy for all the soft biometric classes in our dataset. The accuracy we have achieved with our model is 96.16% for age, 97.74% for ethnicity and 99.2% for age on our UTKface dataset keeping a training loss below 0.1 for the three soft-biometric traits. All the models have been trained into the same environment and it is being uploaded with the source code to the given link below: https://github.com/Kibria10/machine-learning-works.
A deep neural network approach for intelligent crop selection and yield prediction based on 46 parameters for agricultural zone-28 in Bangladesh
(BRAC University, 2018) Islam, Tanhim; Chisty, Tanjir Alam; Roy, Prova; Chakrabarty, Amitabha
Agriculture is the essential ingredients to mankind which is a major source of livelihood and that provides the wide-reaching scope of working opportunities for rural people in underdeveloped or developing countries. Agriculture work in Bangladesh mostly done with old ways which directly affects our economy. In addition to, institutions of agriculture are working with manual data which cannot provide a proper solution for crop selection and yield prediction. The contribution of our thesis is to achieve the best crop selection and yield prediction in minimum cost and effort. Artificial Neural Network considered as robust tools for modeling and prediction. This algorithm aims to get better output and prediction. As well as, support vector machine, Logistic Regression, and random forest algorithm are also being considered in this thesis for comparing the accuracy and error rate. Moreover, all of these algorithms used here just to see how well they performed for a dataset which is over 0.3 million. We have collected 46 features such as – maximum and minimum temperature, average rainfall, types of land, types of chemical fertilizer, types of soil, soil moisture,soil moisture, soil consistency, soil reaction and soil texture and created our dataset for applying into this prediction process. The dataset we have considered are from past ten years (2008-2017) of Bangladesh. Therefore, based on this parameter we will predict the best possible crop selection and yield prediction intelligently.
A federated learning approach for detecting Parkinson’s disease through privacy preserving by blockchain
(BRAC University, 2022-05) Dipro, Sumit Howlader; Islam, Mynul; Nahian, Md.Abdullah Al; Azad, Moonami Sharmita; Chakrabarty, Amitabha; Reza, Md. Tanzim
Parkinson’s disease is a degenerative ailment caused by the loss of nerve cells in the brain region known as the Substantia Nigra, which governs movement. These nerve cells die or deteriorate, rendering them unable to produce an essential neurotransmitter called dopamine. The loss of dopamine in the basal ganglia precludes normal function when the substantia nigra neurons are harmed in large numbers. This results in the motor symptoms of Parkinson’s disease, including tremor, rigidity, decreased balance, and lack of spontaneous movement. For the detection of PD, traditional machine learning algorithms have been used in many research papers. However, traditional ML algorithms always put a risk on the sensitivity of patients’ data privacy. This research proposes a novel approach to detect PD by preserving privacy and security through Blockchain-based Federated Learning. FL may train a single algorithm across numerous decentralized local servers as an improved version of the ML approach instead of trading gradient information. Blockchain can be effectively used to preserve privacy and secure transactions (i.e., gradient) between local and central servers. The proposed model has been tested and evaluated by using three CNN models (VGG19, VGG16 & InceptionV3) in this research, and within these models VGG19 has the best accuracy of 97%. The result demonstrates that this model is very accurate for detecting PD by preserving one’s privacy and security through Blockchain-based Federated Learning.
A hybrid deep learning model for potato leaf disease detection using CNN and vision transformer
(BRAC University, 2024) Saeed, Abdullah; Azad, Afra Binte; Rifat, Md. Abdullah Al; Akter, Orin; Chakrabarty, Amitabha
Finding diseases in potato leaves is important for increasing food yields and lowering losses in agriculture. We describe a strong system that uses the best features of both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to correctly find and group diseases in potato leaves in this study. By combining these two models, our system can successfully pick up both simple and complex patterns in leaf images, which makes it very good at finding diseases. A large set of pictures of potato leaves, including both healthy ones and ones with diseases like Late Blight and Early Blight, were used to train and test the model. Overall, our hybrid model was 97 percent accurate, and its precision, recall, and F1 scores showed that it was good at telling the difference between healthy leaves and those that had diseases. In particular, the model did a great job of finding Late Blight (0.90) and Early Blight (0.96) and it did even better with healthy leaves (0.98). The accuracy of our method is shown by these exact results. We made this answer available by using Flask to create an easy-to-use web application. The app lets users share pictures of potato leaves so that diseases can be found in real time. It gives farmers and agricultural experts immediate feedback, thorough disease information, and treatment suggestions, making it a useful tool for keeping crops healthy. The goal of this study is to connect research with real-life use by combining deep learning methods with useful, scalable solutions. Our work adds to precision agriculture by giving farmers a reliable and effective way to find potato leaf diseases. This leads to better crop management and higher yields.
A localized adaptive architecture for dynamic wireless sensor networks
(BRAC University, 2020-10) Hussain, Sajjd; Ahmed, Muntasir; Joy, Kishanu Roy; Hasan, Md. Mehedi; Chakrabarty, Amitabha
IoT is seen as the next big thing in the world to come. Previously people were less connected with the internet. But in recent years, IoT has seen substantial research and innovation in many technology areas, and uses of the internet are increasing day by day. In the near foreseeable future, the vast number of devices connected together will result in highly complex networks that would cause an exponential increase in computation times, latency and power consumption. In this paper, we take a modular approach to dealing with the issue. We propose a Weight-based Adaptive Neighbor Localization (WANLoc) Algorithm that creates a highly opti- mized network based on localized network architecture that is network independent. This means that the sensor nodes in this network do not need to communicate with any cluster head or sink in order to function. In addition, WANLoc is intended to function in a dynamic wireless sensor network where nodes enter and exit the network regularly. This also means that it is exceptionally good at handling node failures. Nodes in this network only communicate with their immediate neighbors, maintained in a set of dynamic tables designed to reduce lookup times when deciding whom to forward data to. This also means that nodes can now be configured with minimum transmission capabilities, saving a lot of precious energy in the process.
A machine learning approach on classifying orthopedic patients based on their biomechanical features
(BRAC University, 2018-04) Hasan, Kamrul; Islam, Safkat; Samio, Md. Mehfil Rashid Khan; Chakrabarty, Amitabha
A person’s orthopedic health condition can be detected from his biomechanical features. Application of machine learning algorithms in medical science is not new. Different algorithms are applied to detect diseases and classify patients accordingly. This paper aims to assist specialists to predict the type of orthopedic disease. In this paper we have applied various machine learning algorithms to find out which one works most accurately to detect and classify orthopedic patients. Each of the patients in the dataset is represented by six biomechanical attributes derived from the shape and orientation of pelvis and lumbar spine. We performed our operation in two stages and got an average accuracy of more than 90 percent for most of the algorithms, whereas Decision Tree (DT) algorithm stood out from the rest providing 99% accuracy.
A new approach for processing X-ray data for detecting mini-halos in galaxy clusters
(BRAC University, 2021-06) Jafar, Zareef; Mim, Armin Nahar; Roy, Sworna; Chakrabarty, Amitabha; Alam, Md. Golam Rabiul
Galaxy clusters, which may contain hundreds or thousands of galaxies, are the biggest gravitationally bound objects in the universe. These clusters are filled with heated X-ray emitting plasma. Radio mini-halos are radio sources which are found in some cool-core non-major merging galaxy clusters and study of these mini halos helped scientists to unveil many other mysteries of how the universe works. Till now 23 such galaxy clusters have been found, many with confirmed minihalo and some with potential mini-halos. Even a few years ago observational data of this kind of astronomical observations were available only to the astronomical communities and specific research institutions. As more and more data is publicly available it made a way for Amateur astronomers and students who want to work with real data. In our work we use such publicly accessible X-ray data of these galaxy clusters from Chandra X-ray Observatory for determining the cold front which ultimately will give us the location of these mini halos. The main goal of our thesis is to develop a workflow using some open-source Astro data processing packages and including our own python script for such purpose. These software tools include CIAO, ClusterPyXT pipeline, conda and our own python script. This workflow includes setting up the proper software environments, list of packages needed, rules for choosing different input parameters for downloading as well as processing data successfully and some sources to compare the generated output. From start to end this process can be successfully run on any moderate modern desktop computer.
A new routing algorithm for symmetric rearrangeable networks and emerging applications
(© 2012 IEEE, 2012) Chakrabarty, Amitabha; Collier, Martin J.
An excavation tele-robotic system is developed to excavate and collect lunar regolith. The excavator has been developed by the team BRACU ChondroBot consisting students from BRAC University for NASA's 2nd Annual Lunabotics Mining Competition (LMC) 2011. Considering the requirement of NASA and calculating the load, friction and power the mechanical excavator was designed and built. The dimension of the excavator is 1.45m ×.74m × 1.48m and the weight is 80 kg. It consists of two excavation arm and one pulley system depositor bucket. Two conveyer belt type wheels are used to drive the robot. Efficient control and communication is always a big challenge for a Tele-robot. In our developed system hardware can be operated both manually and remotely through a web browser by logging in from any computer without direct visual and auditory access to the hardware. A unique control circuit, graphical user interface and communication module for two terminals are also developed for remote access.
A new routing algorithm for symmetric rearrangeable networks and emerging applications
(© 2012 IEEE, 2012) Chakrabarty, Amitabha; Collier, Martin J.
An excavation tele-robotic system is developed to excavate and collect lunar regolith. The excavator has been developed by the team BRACU ChondroBot consisting students from BRAC University for NASA's 2nd Annual Lunabotics Mining Competition (LMC) 2011. Considering the requirement of NASA and calculating the load, friction and power the mechanical excavator was designed and built. The dimension of the excavator is 1.45m ×.74m × 1.48m and the weight is 80 kg. It consists of two excavation arm and one pulley system depositor bucket. Two conveyer belt type wheels are used to drive the robot. Efficient control and communication is always a big challenge for a Tele-robot. In our developed system hardware can be operated both manually and remotely through a web browser by logging in from any computer without direct visual and auditory access to the hardware. A unique control circuit, graphical user interface and communication module for two terminals are also developed for remote access.
A practical approach to convert silo web application into multi-tenant cloud-based SaaS application
(BRAC University, 2025-10) Khan, Omar Hassan; Chakrabarty, Amitabha
Generally, SME software firms start developing an application with specific require- ments and intended to use it as Application Service Provider (ASP). With no argu- ment ASP fits best for agile development and meets development deadlines. In ASP model, we deploy the system and cater update, customization and troubleshooting by ownership. When we have the application deployed for more than a handful of users then it gets difficult to maintain the application for each owner. In Software as a Service (SaaS) model service is provided to each subscriber individually but under- neath the application is maintained as one single instance. Converting ASP model applications to SaaS model application helps the provider to reduce development and operational costs, effortlessly update and do bug fixes, and attain troubleshooting effectively. The conversion of ASP to SaaS is not straightforward. The conver- sion starts with separation common modules and subscriber specific modules. Then there is existing database customization along with master table implementation for the subscription. Finally, we need to select proper cloud-based IaaS with different services to obtain maximum usage with minimal billing. This project address the challenges of converting an ASP to a multi-tenant SaaS model with scalability, cost efficiency and ease of maintenance. Using the cloud as infrastructure gives the advantage of high security, reduces overhead cost and streamlines integration and deployment. This paper aims to provide a structured implementation technique to convert an existing singleton web application to SaaS on multi-tenant shared infrastructure model. These improvements help software firms achieve better operational efficiency while offering a more sustainable SaaS solution. Also helpful for the researchers and engineers to easily implement multi-tenant architecture in action.
A sentiment-based comprehensive rating model with extensive dataset for nationwide hospital rating in Bangladesh using natural language processing and machine learning
(BRAC University, 2025-10) Auishik, Asiful Kanzan; Al-Zahir, Fariha Mohammed; Samia, Magferah Sultana; Hossain, Moumita; Chakrabarty, Amitabha
Bangladesh has a huge number of public and private hospitals in its districts. However, there is no trustworthy, unbiased resource for patients to choose the best hospitals concerning quality of service. The existing star-based rating systems are prone to manipulation and do not capture detailed feedback from the patients. This paper proposes an advanced model of hospital rating that grades the hospitals based on online reviews of the patients, considering aspects like patient experience and quality of care. This proposed model employs NLP and ML techniques to analyze the sentiment of patient feedback and extract insights from it. It is expected to provide a data-driven hospital rating system based solely on user experiences by integrating various dimensions of hospital service quality to identify strengths and areas for improvement across the country. A large dataset of structured and unstructured reviews is collected from online platforms. Text mining and advanced NLP techniques process sentiment data. Various machine learning models, such as SVM, BERT, and CNN, are trained and validated on pre-processed data for sentiment prediction. Steps to achieving this objective involve data collection, data preprocessing, sentiment analysis, and, eventually, aspect-based sentiment analysis using zero-shot aspect detection to generate hospital ratings based on 4 aspects: treatment quality, cleanliness, affordability, and service quality. Rating generation classify sentiments as positive, negative, or neutral, thereby dynamically, and in real time, rating a hospital. This model provides a more reliable and nuanced rating system that allows for transparent comparisons across hospitals and actionable insights into strengths and weaknesses. The framework is adaptable to other sectors, such as education and retail, providing an enlarged scope of application for sentiment analysis in service quality evaluation. For sentiment prediction of the reviews, BERT proves to be the best performing model out of all the models in terms of accuracy, precision, recall, and F1 score, producing 94.1%, 94.7%, 94.1%, and 94.4% respectively. For the aspect-based ranking of hospitals, the model is most confident in detecting treatment quality as it produces the highest teacher threshold of 0.51 for this aspect. It also produces the highest precision, F1 score, and agreement of 0.98, 0.97, and 0.96, respectively, for treatment quality, whereas, both affordability and service quality score the highest recall of 0.99.
A supervised learning approach by machine learning and deep learning algorithms to predict type II DM risk
(BRAC University, 2019-09) Farabe, Abdullah Al; Sharika, Tarin Sultana; Raonak, Nahian; Ashraf, Ghalib; Chakrabarty, Amitabha
The application of Arti cial intelligence (AI) has become a valuable part of medical research. These days diabetes is one of the top maladies on the planet. Nowadays it has become a common disease and alarming as people are living in polluted areas and eating unhygienic foods. People with diabetes are probably going to pass on at a more youthful age than individuals who don't have diabetes. We hope this study could be very helpful in medical science to predict the risk score of type II Diabetes Mellitus (DM). Our model consists of four machine learning algorithms which are- K-Nearest Neighbor, Random forest, Decision tree and Logistic Regression. These algorithms have been applied on a dataset containing 15000 type 2 diabetes patients along with eight features that describe the state of patients such as glucose, BMI, age, pregnancy, blood pressure (BP), Diabetes Pedigree Function, Skin thickness and insulin. Moreover, one deep learning algorithm called CNN has been used. All of the ve algorithms have been used on the dataset and the Random forest gives the best accuracy of almost 92.60 percent where other algorithms give less accuracy.
A supervised machine learning approach to predict vulnerability to drug addiction
(BRAC University, 2019-08) Faisal, Fahim; Shahriar, Arif; Mahmud, Sohan Uddin; Shuvo, Rakibul Alam; Chakrabarty, Amitabha
There are signi cant amount of di erences between an addicted and non-addicted person on their social and familial behavior. In our thesis we tried to nd out the characteristics of a person related to his social and familial life and also health issues that can prove his vulnerability to drug addiction. The research was held on the context of the people of Dhaka, Bangladesh and on an age group of 15 to 40 years. A primary data set was constructed which include 498 samples. For constructing the questionnaire Addiction Severity Index and WHO's Assist Scale were followed along with the help of psychologists and specialists on drug addiction. For addicted person's data we reached some rehabilitation center of Dhaka and for non-addicted person's data we communicated di erent aged group people of di erent colleges and universities. 498 samples where one sample consisted of 60 features were trained and tested by supervised machine learning approach. Reliability of the data set was validated by Cronbach's Alpha Nominal Test. 10 algorithms were incorporated including Neural Network, Deep Belief Network, Random Forest, XGBooster etc. and their results were compared. Among the algorithms, XGB came up with the highest number of accuracy of 95.20% and KNN delivered the least which is 88.97%. In order to select important features mRMR, Chi-square, Principle Component Analysis techniques were used. From feature selection we got the key features of an addicted person's behavior that were in uential for their drug abuse. This will help people to understand if a person is going to be vulnerable to addiction or not based on their health issues and social and familial behavior.
Adaptive traffic signal control for urban intersections using reinforcement learning: a SUMO simulation case study on Gulshan-2
(BRAC University, 2025-02) Shamim, MD Shahadat Hossain; Chowdhury, Tawhid; Deep, Sadid Arman; Bhuban, Riazul Hoque; Chakrabarty, Amitabha
Traffic congestion is an ever-growing concern in rapidly increasing population and increase of vehicles on roads. On top of that, our road condition, pedestrian be- havior, and driving behavior makes the situation much worse. As Bangladesh has started adapting to the fixed-time traffic light system, finding the optimal adaptive traffic light management system is required to mitigate the issue. Keeping the lack of intra-vehicular communication devices in mind, we have studied different traffic light controlling systems and how these perform on Bangladeshi roadways. In our study, we have used traffic simulator to simulate Gulshan-2 intersection, one of the major congestion points in Dhaka city. Based on traffic data, we have created a vehicular network system to find out how the existing traffic light models per- form on mitigating the traffic jam using an adaptive traffic management system. In our study, we simulate the Gulshan-2 intersection using a realistic traffic dataset de- rived from field data and analyze three control models, Fixed-Time, Q-Learning, and Deep Q-Network (DQN). The results show that in ideal traffic condition, when no anomalies are present; the Q-Learning controller reduced average vehicle delay by approximately 60-70% and improved throughput by 35-45% compared to the fixed- time system, demonstrating its superior adaptability to dynamic traffic flows. How- ever, when we add real-life anomalies such as potholes and jaywalking, the baseline controllers’ performance drops. By acknowledging the drawbacks of QL controller in abnormal scenarios, we have introduced a hybrid model CRQL (Congestion Re- sponsive Queue Learning) and made some improvements. In abnormal situation CRQL performs better in regard to throughput, which is a 76.9% improvement over Fixed-Time, 37.4% improvement over Q-Learning and 215.2% improvement over DQN. This reflects a overall performance improvement under disruptive and non- stationary environment. In short, the study helps to see how existing models perform under real-life anoma- lies, which factors are significant to create Ad hoc model for our extreme pedestrian patterns and road conditions and how our proposed solution; the CRQL ad hoc or hybrid model performs to address the problems.
Advanced machine learning techniques for personalized alopecia areata intervention modeling
(BRAC University, 2025-01) SHOWMIK, MUTASIM FOUAD; RUHAN, MOHAMMED; ISLAM, MD MOZAHIDUL; SULTANA, RAFIA; RISHTA, SADIA ZABIN; Chakrabarty, Amitabha
Alopecia Areata (AA) is a skin condition that causes hair loss, ranging from small patches to complete baldness. Diagnosis and treatment are not very easy because of the complex interaction among genetic, autoimmune, hormonal, environmental, and sociodemographics. In this study, machine learning techniques were explored to assess different stages of Alopecia Areata to allow early diagnosis and treatment. Three complementary datasets consisting of clinical, genetic, environmental, and demographic features were used to build and validate predictive models. Several machine learning models were trained, including Meta Classifiers, K-nearest neigh- bors (KNNs), and Random Forest models. Among all the models tested, the Meta Classifier achieved the highest accuracy and was selected as the base model for fea- ture extraction by LIME and SHAP. Identified key features impacting the models were fed into the fusion datasets for having advanced models such as 1D CNN with autoencoder. The best-performing model was the 1D CNN with an autoencoder, with a 96.78% accuracy, able to integrate the features extracted. This study shows the capability of machine learning to aid early diagnosis and personalized manage- ment of alopecia areata and subsequently provides an avenue for better and more reliable intervention strategies.
Advanced task scheduling algorithm for IoT Based FOG communication model
(BRAC University, 2020-04) Mohammed, Zaber; Chowdhury, Riham; Rozario, Stanley Dip; Sakin, Sayed Bin Amirul; Chakrabarty, Amitabha
Internet of Things (IoT) is hugely dependent on Cloud Computing. Cloud comput- ing uses a high degree of polymerization calculation mode and so it cannot ensure e ective use of resources like computing, storage, etc. FOG computing is a devel- oping paradigm that broadens computation, communication and storage facilities towards the edge of a network. It is used to improve e ciency along with the re- duction of transmitted data for processing to the cloud. Although the primary aim of FOG computing is to improve the processing speed of cloud computation, it has many challenges such as task scheduling, resource allocation, security, etc. Among these challenges handling incoming requests to improve latency and throughput is one of the crucial factors. As a solution, the proposed model uses a multi-layered FOG model in which tasks are scheduled on the basis of priority based on the re- quest type to increase the e ciency of the current FOG model. Firstly, the proposed model creates a rule list based on the user's request priority. While creating the rule list the model will use advance caching mechanism based on the request type in the di erent layers to improve latency and throughput. When a user sends data to the FOG, it nds its con gured layer of the FOG cloud on which the data will be processed. Stored data will be loaded in the corresponding layer based on packets' priority to make the computation faster. The proposed model has been simulated using Microsoft Azure. In the simulation, the inbound data after caching was more than 70 MB per 30 seconds wherein the traditional cloud, the inbound data rate was around 30 MB per 30 seconds. Therefore, after caching the data, the model performed twice faster than the traditional cloud.