1. Real-Time Tele-Consultation for Cervical Cancer Screening Using a Machine Learning-Enabled Mobile Application
Collaborating Institution: Uganda Cancer Institute
This project developed a real-time tele-consultation mobile application to support frontline healthcare workers conducting cervical cancer screening using Visual Inspection with Acetic Acid (VIA) in resource-limited settings. In response to the high burden of cervical cancer in Uganda and limited access to specialist care, the system builds on a mobile colposcopy initiative deployed at the Mayuge satellite clinic by the Uganda Cancer Institute. The Android-based application enables healthcare workers to capture and upload cervical images, which are immediately reviewed by remote gynaecologists who provide real-time diagnostic impressions and guidance. The platform securely stores patient data to ensure continuity of care and integrates a machine learning model that classifies images as “positive” or “negative,” highlights suspected lesion regions, and supports clinical decision-making when specialists are unavailable. Additional features include automated SMS reminders for follow-up screening and scalable deployment on Android smartphones, ultimately improving access to expert consultation, enhancing diagnostic accuracy, and strengthening follow-up mechanisms in underserved communities.
2. Development of Machine Learning Datasets for Crop Pest and Disease Diagnosis Using Imagery and Spectrometry Data (non-funded)
This project focuses on advancing crop pest and disease diagnosis from traditional visual inspection to data-driven approaches using machine learning and computer vision. While smallholder farmers and extension workers collect geo-coded, time-stamped field data via mobile applications for crops such as beans, cassava, bananas, and tomatoes, existing image datasets have not been sufficiently curated or standardized for broader machine learning applications. Additionally, by the time visual symptoms appear, crop damage is often advanced, limiting intervention effectiveness. This work involved curating and preparing crop disease image datasets, integrating leaf spectral data for early disease detection—potentially up to six weeks before visible symptoms—and investigating plant metabolite changes under real field conditions rather than controlled environments. By transitioning disease detection methods to reproducible field settings and improving dataset quality, the project contributes to early disease identification, improved yield protection, and the development of open, high-quality resources for the machine learning research community.
3. Design of a Machine Learning-Based Traffic Control System
Study Location: Wandegeya Junction, Kampala, Uganda
Collaborating Authority: Kampala Capital City Authority
This project addressed severe traffic congestion at Wandegeya Junction in Kampala by developing a reinforcement learning-based adaptive traffic control system to replace inefficient fixed-time traffic signals. Using manually collected vehicle counts and historical traffic data from the Kampala Capital City Authority, the system applied Q-learning, a model-free reinforcement learning algorithm enhanced with neural networks, to dynamically adjust traffic signal durations based on real-time traffic conditions. The model was evaluated using Simulation of Urban Mobility (SUMO) within a simulated vehicular network environment, where results demonstrated improved traffic flow efficiency compared to conventional signal control methods. The project highlights the feasibility of AI-driven adaptive traffic management systems in rapidly growing urban African contexts.
4. Deep Learning-Based Pothole Detection System (non-funded)
This project developed a computer vision-based pothole detection system aimed at addressing road safety risks, vehicle damage, flooding, and traffic congestion caused by delayed infrastructure maintenance. A dataset of 836 road images was collected from smartphone video recordings captured inside a moving vehicle, after which images were resized, annotated with bounding boxes, and converted into COCO format for model training. The EfficientDet object detection algorithm was implemented and optimized to detect potholes under varying road conditions, achieving an Average Precision (AP) score of 0.525. The project resulted in a trained pothole detection model, a structured and labeled pothole image dataset, and an automated detection pipeline capable of supporting faster road condition assessment and smart city infrastructure maintenance systems.