Student : MMJ Mphekwane
About the student
My name is Modjadji Mphekwane, and I'm currently a BSc Honours student in Computer Science at the University of Johannesburg, with a specialisation in Artificial Intelligence. This follows my undergraduate degree in Computer Science and Mathematics at the University of the Witwatersrand.I have a deep passion for implementing technology that can solve real-world problems, which is precisely why I chose to specialize in AI for my Honours degree.
About the Project
This research project, "AI-Driven Mammography Analysis," investigates developing an efficient, scalable diagnostic tool to overcome critical detection constraints (like the shortage of radiologists and delayed diagnosis) in the South African public sector. The primary aim is to improve the accuracy and speed of identifying breast lesions. The core work involved designing and testing a two-stage AI pipeline built for computational efficiency. The architecture combined a U-Net model for precise pixel-level lesion segmentation (Stage 1) with a Random Forest classifier (Stage 2) trained on extracted features and mammogram metadata. The study successfully demonstrated the technical feasibility of this low-resource approach. However, the evaluation revealed significant performance failures. The U-Net segmentation model was fundamentally inaccurate, achieving a critically low Dice Score of only 32%, severely limiting its ability to localize lesions. Consequently, the Random Forest classifier suffered, showing an inability to classify high-risk morphologies (like Architectural Distortion). Most critically, the system only achieved a 50% Malignant Recall, meaning the model failed to identify half of all true cancer cases, revealing an unacceptable safety gap. The project concludes with a detailed roadmap for future work, prioritizing advanced deep learning techniques to close this gap. This includes implementing a dual-phased CNN for classification and redesigning the detection stage to meet a clinically safe target of achieving over 90% Malignant Recall.
