(MSC Thesis) Containerized edge-cloud service orchestration for medical image computing [2025]
Author: Abu Taher
Abstract: Traditional standalone medical devices perform two critical functions: acquiring data from a patient's body using various sensors and processing this raw data using an algorithm. The algorithm requires a runtime environment to execute, typically a dedicated processing unit, such as a workstation. Standalone devices' isolated architectures limit the deployment of real-time applications. This architecture typically lacks computational power, scalability, mobility, resiliency, and the benefits of multi-tenant shared resource utilization. To mitigate these limitations, the processing functionality would be optimal when abstracted into containerized microservices and distributed across an edge–cloud continuum. The continuum consists of a three-tier heterogeneous architecture, where upper tiers provide increasingly higher computational capacity, while lower tiers are closer to the users and data sources, leading to lower latency, among other benefits. Orchestrating microservices in the edge-cloud continuum further introduces new challenges, such as task offloading, task scheduling, and load balancing.
To address the aforementioned challenges, this thesis presents three main contributions. First, a multi-cluster, geo-distributed federated orchestration framework is developed for orchestrating deep-learning-based denoising services across the edge-cloud continuum. This framework can also be used to deploy any other distributed application. The framework is designed to offer scalability, resiliency, a multi-tenant shared resource utilization mechanism, and the capability to enable mobility-aware orchestration. Second, a producer-worker algorithm is presented for efficiently handling distributed task offloading, task scheduling, and load balancing across the edge-cloud continuum. Finally, an integrated monitoring service is implemented to provide insight into system performance, resource utilization, and cluster behavior.
The multi-cluster orchestration framework and the microservices are developed and evaluated separately. The experimental data show that the producer-worker model can efficiently handle distributed task offloading, task scheduling, and load balancing across the edge-cloud continuum.
In conclusion, the designed multi-cluster orchestration framework and the experimentally validated microservices provide a pathway toward mitigating both the limitations of traditional standalone medical devices and the coordination challenges of multi-tier orchestration. Integrating the containerized microservices with the federated multi-cluster orchestration framework and mobility-aware autonomous orchestration will be addressed in future work.
Cotton Leaves Disease Classification Using Custom Convolutional Neural Network (CNN) Model [ 2023 ]
Author: Abu Taher, Md. Foisal Hossain
Abstract: Cotton is a major industrial crop for Bangladesh's economy. Bangladesh's textile sector generates 27% of the country's GDP. 95% of the cotton used in the textile sector is imported, which lowers the gross profit margin. The biggest barrier to reaching high production is a variety of leaf diseases. To guarantee good cotton production, early disease identification and the correct insecticide are essential. The classification of images using CNN models is particularly efficient and effective. In this study, we created a unique CNN model to identify and categorize five illnesses of cotton leaves. The model consists of five convolutional layer and four dense layers. The model has simple linear architecture. Both augmented and non-augmented data were used to evaluate this model. We used five different types of photos of cotton disease to train the model. For augmented and non-augmented data, our model produced accuracy levels of 98.5% and 94.64%, respectively.
Author: Rabiul Hasan, Shah Muhammad Azmatullah, Avizit Nandi, Abu Taher
Abstract: The lung infection known as pneumonia is commonly carried on by fungal infections, bacterial infections, or virus infections. Particularly in undeveloped and developing countries where pollution rates are very high, people live in unhygienic circumstances, overpopulation is common, and there is insufficient healthcare infrastructure, it affects a large number of people. Therefore, it is crucial to find pneumonia early in order to ensure proper treatment and increase survival rates. A chest X-ray examination is the most popular method for diagnosing pneumonia. The examination of CXRs is difficult and vulnerable to subjectivity. In our research, we created a system for automatically detecting pneumonia from CXR images. To deal with the shortage of data, we applied deep transfer learning and created an ensemble using three CNN models: VGG16, MobileNetV2 and DenseNet169. The weighted average ensemble method was used for this purpose. The suggested method successfully identified cases of pneumonia and normal cases with recall rates of 93% and 89%, maximum accuracy of 92%, and precision rates of 93% and 89%. Our approach outperformed the widely used ensemble techniques, and the results were better than those of existing methods.