A Ugandan Artificial Intelligence researcher has developed a new AI-powered medical imaging model that could improve how diseases such as tuberculosis and brain tumours are detected, while preserving critical diagnostic information often lost in conventional medical image analysis systems.
Mabirizi Vicent, a computer scientist and Artificial Intelligence researcher at Kabale University, developed the innovation as part of his doctoral research at Mbarara University of Science and Technology (MUST), where he successfully defended his PhD at the age of 29. His research focused on one of the growing challenges in AI-assisted healthcare: how medical imaging systems process and analyze Digital Imaging and Communications in Medicine (DICOM) files, the standard format used in hospitals worldwide to store X-rays, CT scans, ultrasound images, and Magnetic Resonance Imaging (MRI) scans.
According to Mabirizi, most existing Artificial Intelligence systems used in medical imaging first convert DICOM files into standard image formats such as JPEG or PNG before analysis. While this simplifies processing, it often removes critical embedded clinical metadata that may influence diagnosis and interpretation. Research studies have proved that this loss of metadata can introduce classification bias and reduce diagnostic accuracy, particularly in sensitive clinical environments where image quality and patient information are essential for decision-making.
To address this challenge, Mabirizi developed a Convolutional Neural Network (CNN) model capable of directly processing raw DICOM medical files without converting them into other image formats. The system preserves both the image data and the embedded metadata, allowing the AI model to make more informed and accurate classifications.
The model was evaluated using chest X-ray images for tuberculosis detection and MRI scans for brain tumour classification. Results from the study showed strong performance across both imaging applications. In tuberculosis detection using chest X-rays, the model achieved:
92.9% precision , 88.4% recall , 90.6% F1-score , 90.9% accuracy
For brain tumour classification using MRI scans, the model recorded:
80.0% precision , 79.4% recall , 79.7% F1-score , 85.5% accuracy
The findings suggest that preserving diagnostic metadata during AI analysis can significantly improve classification reliability while reducing the risk of bias in automated medical diagnosis systems.
Beyond academic significance, the innovation could have practical implications for healthcare systems, particularly in resource-limited environments where access to medical specialists remains limited and hospitals increasingly rely on digital diagnostic support tools.
Researchers believe such AI-powered systems could support faster clinical decision-making, strengthen disease detection, and improve healthcare delivery in areas facing shortages of radiologists and diagnostic infrastructure.
Mabirizi’s work also reflects the growing contribution of African researchers to global Artificial Intelligence and health technology innovation.
His research has been presented at several academic symposiums and international conferences, including the Deep Learning Indaba conferences held in Dakar, Senegal, and Kigali, Rwanda.
As Artificial Intelligence continues to reshape global healthcare systems, innovations such as Mabirizi’s CNN-based DICOM processing model highlight the increasing role African universities and researchers are playing in advancing locally relevant scientific solutions.
The research also underscores a broader shift taking place within Uganda’s higher education sector, where universities are increasingly contributing to innovation in healthcare technology, data science, and AI-driven research with both regional and global relevance.
