http://hdl.handle.net/123456789/10320 Thu, 25 Jun 2026 14:48:28 GMT 2026-06-25T14:48:28Z http://hdl.handle.net/123456789/21303 CyberSight An AI Powered Diabetic Retinopathy Detection System Ayesha Shamim, 01-131222-011; Muhammad Bilal Masood, 01-131222-030 CyberSight is an AI-powered clinical decision support system developed to assist ophthalmologists and general physicians in the early detection and severity grading of diabetic retinopathy using retinal fundus images. Diabetic retinopathy is one of the leading causes of blindness worldwide, especially among diabetic patients who do not receive timely screening and treatment. As the number of diabetes cases continues to rise globally, healthcare systems face increasing pressure to provide faster and more accessible retinal screening services. In many rural and under-resourced areas, the shortage of trained specialists makes regular eye examinations difficult, which often results in delayed diagnosis and permanent vision damage. CyberSight aims to address this challenge by providing an intelligent and accessible platform that supports doctors in identifying diabetic retinopathy at an earlier stage while ensuring that the final medical decision always remains under professional supervision. The system is designed using a modern web-based architecture that combines a React frontend with a FastAPI backend for smooth communication and efficient request handling. The core AI component is based on a Vision Transformer (ViT) deep learning model trained to classify retinal fundus images into five diabetic retinopathy severity categories, ranging from no diabetic retinopathy to proliferative diabetic retinopathy. Before prediction, CyberSight performs several validation steps to improve reliability and reduce errors during inference. The uploaded image is checked for supported formats such as JPEG and PNG, image quality and resolution are verified, and a CLIP-based zero-shot validation mechanism confirms that the uploaded image is actually a retinal fundus image. These validation measures help prevent unrelated or poor-quality images from affecting prediction accuracy and improve the overall robustness of the system. In addition to accurate prediction, CyberSight also focuses on transparency and explainability, which are important challenges in medical AI systems. Many deep learning models are often criticized for behaving like “black boxes” because doctors cannot easily understand how predictions are generated. To solve this issue, CyberSight integrates an occlusion-based explainable AI approach that generates visual heatmaps highlighting the image regions that most influenced the model’s decision. This allows doctors to better interpret and verify AI predictions, improving trust and usability in real clinical environments. The platform also supports automatic diagnostic report generation that includes confidence scores, severity explanations, and medical disclaimers. These reports simplify communication between doctors and patients and make clinical documentation easier. Overall, CyberSight demonstrates how artificial intelligence can be integrated into healthcare systems as a supportive tool to improve early detection, enhance screening efficiency, and assist medical professionals in providing better patient care. Supervised by Engr. Rafia Hassan Thu, 01 Jan 2026 00:00:00 GMT http://hdl.handle.net/123456789/21303 2026-01-01T00:00:00Z http://hdl.handle.net/123456789/21302 Smartlab: Next-Gen Virtual Science Laboratory (Student Level) Abdi Fatah Yusuf Mohamed, 01-131222-002; Rumman Saeed, 01-131222-055 Hands on laboratories are very critical in the study of Physics and Chemistry concepts of science. Nevertheless, most schools encounter issues like lack of resources in the laboratories, lack of equipments, safety issues and lack of accessibility to laboratory physical set-ups. As a solution to the problems stated above, this project introduces SmartLab: AI-Powered Virtual Lab, a web-based solution, that aims to offer students an interactive and intelligent virtual laboratory experience that complies with the FBISE curriculum. The SmartLab system allows students to conduct virtual experiments in an easy to use interface through which students can choose experiments, change the experiment parameters, simulate the results of the experiment and create reports about the experiment. The system has been equipped with a machine learning model that offers the predictive-feedback and helps a learner to interpret the expected outcomes of experiments. This has been a brilliant feature that boosts conceptual learning as it directs the students when running experiments and it minimizes the errors encountered when choosing parameters. The system is built on the basis of recent web technologies such as Next.js, React and TypeScript, PostgreSQL, and Prisma ORM that guarantee scalability, maintainability, and effective data management. The modular design couples the frontend elements, backend services as well as database functions and enables unproblematic communication between modules of the system and future-proofing of expansion into multiple experiments. Moreover, the platform has role-based access control between students, teachers and administrators, which allows monitoring of the experiments, tracking performance and management of the system in a systematic way. SmartLab offers a convenient, risk-free and economical alternative to a conventional lab space by using simulative experimentation in combination with assistance of artificial intelligence. The proposed system on the whole enhances the teaching of science by increasing student engagement, remote learning, and also offers a smart digital lab to conduct Physics and Chemistry practical experiments. Supervised by Engr. Sulman Zafar Thu, 01 Jan 2026 00:00:00 GMT http://hdl.handle.net/123456789/21302 2026-01-01T00:00:00Z http://hdl.handle.net/123456789/21301 CorneaInsight ( Insightful Analytics for Early Keratoconus Detection) Adil Shahzad, 01-131222-007; Muhammad Usman Ali, 01-131222-038 CorneaInsight (Insightful Analytics for early keratoconus Detection),, gives a reliable tool to ophthalmologists for detecting SKC in order to avoid the development of such a dangerous eye condition as keratoconus. In particular, it is really vital to diagnose SKC timely and advise patients on how to protect their health before any interventions with their eyes take place. In other words, KEDS can be considered a "smart second opinion." In fact, this system takes into account two aspects at once, namely objective information (for example, specific measurements of the eye) and subjective information (eye images). The algorithm works by using a two-stage AI approach that consists of two methodologies in order to provide more accurate results. Firstly, the methodology referred to as EfficientNet-B0 is focused on pattern recognition within eye images, whereas the other part of the model is called ExtraTreesClassifier and focuses on tabular data processing. Integrating these two methodologies into one framework helps to produce a single outcome on whether the eyes in question are healthy or are diagnosed with Keratoconus. In order to make healthcare professionals trust the outcome of the predictions, the proposed algorithm provides the reasoning behind the decisions made. To bridge the gap between complex algorithmic processing and clinical utility, the framework integrates dedicated Explainable AI (XAI) mechanisms. This ensuring that the system does not operate as a standard "black box," but instead generates transparent decision pathways alongside its classification. By exposing the explicit features and visual markers driving the diagnostic output, the algorithm provides medical professionals with the necessary interpretability to confidently audit, validate, and trust the system's conclusions in a clinical setting Supervised by Engr. Aamir Sohail Thu, 01 Jan 2026 00:00:00 GMT http://hdl.handle.net/123456789/21301 2026-01-01T00:00:00Z http://hdl.handle.net/123456789/21299 Sonar Shield Muhammad Arqam, 01-131222-028; Zeeshan Ahmad Abbasi, 01-131222-051 Sonar Shield is an intelligent dual module AI solution developed with the primary objective of solving one of the key problems of contemporary signal processing technologies that relates to the loss of valuable information due to noise in two unique yet equally essential environments: under-water acoustic environment and satellite communication. Two unique modules are incorporated into one universal platform, which can be used by scientists, engineers, and specialists working in the corresponding fields of expertise. The Aqua Noise Reduction Module addresses underwater communication interference from sources like ocean currents, marine life, and vessel vibrations. It utilizes a CNN-based classifier to identify specific noise types within a signal, then automatically selects the most effective deep learning architecture such as Autoencoders, U-Nets, and WaveNet to perform denoising. The resulting highquality audio is then objectively validated using SNR, PESQ, and STOI performance metrics. The second module is the Satellite Signal Noise Reduction Module. The problem addressed in this module is that of error propagation within long-distance satellite communication that can be affected by environmental noise, electromagnetic interferences, and degraded communication channels. To solve this problem, BiGru Decoder sequence models along with FEC, namely Reed-Solomon coding algorithm, will be used to detect and correct errors in telemetry data obtained from satellite signals. The Sonar Shield project is a modular and scalable signal enhancement framework built using Python, TensorFlow, and PyTorch[3]. It integrates a user-friendly interface with advanced data visualization and automated reporting, providing a cross-platform solution for Windows, Linux, and macOS that demonstrates the practical effectiveness of AI in signal processing. Supervised by Dr. Joddat Fatima Thu, 01 Jan 2026 00:00:00 GMT http://hdl.handle.net/123456789/21299 2026-01-01T00:00:00Z