Victoria University Study Finds Lightweight AI Model VGG16 Excels in Chest X-Ray TB Diagnosis

In a landmark development that could reshape tuberculosis screening in low-resource settings, researchers from Victoria University, Makerere University, and the University of Cape Town have unveiled compelling new evidence showing that a lightweight deep learning model, VGG16, significantly outperforms more complex neural network architectures in detecting TB from chest X-rays. The peer-reviewed findings, published in the July 2025 edition of JMIRx Med, reveal that VGG16 achieved a staggering 99.4% classification accuracy—surpassing heavyweight models such as ResNet152 and Inception-ResNet-V2, despite requiring far less computational power.

Spearheaded by Dr. Lillian Tamale, Head of the Department of Computing and Information Science at Victoria University, the study represents a major step forward in the quest for scalable, affordable diagnostic solutions in regions hardest hit by tuberculosis. The project was conducted in collaboration with Mr. Alex Mirugwe of Makerere University and Dr. Juwa Nyirenda from the University of Cape Town. Together, the team set out to investigate whether lightweight neural networks could provide comparable or superior diagnostic accuracy without the need for expensive infrastructure.

Using a dataset of 4,200 chest X-ray images—comprising 700 TB-positive cases and 3,500 normal scans—the researchers evaluated six convolutional neural network (CNN) architectures: VGG16, VGG19, ResNet50, ResNet101, ResNet152, and Inception-ResNet-V2. The models were benchmarked across standard performance metrics including accuracy, precision, recall, F1-score, and AUC-ROC, along with training time and computational efficiency. VGG16 emerged as the top performer across nearly every category, offering a rare combination of speed, accuracy, and resource-friendliness.

“This is a game-changer for AI in public health,” said Dr. Tamale in an interview following the publication. “We’ve demonstrated that highly effective diagnostic tools don’t have to come with high hardware demands. With VGG16, even rural clinics operating on minimal computational infrastructure can access accurate, AI-assisted TB screening. This democratizes the technology in a meaningful way.”

One of the study’s more surprising findings challenged long-held practices in deep learning: the use of data augmentation. The researchers applied common augmentation techniques such as image rotation and flipping in an effort to enhance the training dataset. However, these modifications failed to improve model performance, suggesting the original dataset was already sufficiently diverse. The insight could influence future design approaches for AI models in medical imaging, particularly when operating under tight computational or data constraints.

The study arrives at a critical juncture in global health. According to the World Health Organization’s 2023 report, tuberculosis remains among the leading causes of death from infectious diseases, with over 10 million new infections and 1.5 million fatalities recorded annually. Yet in many low- and middle-income countries, traditional diagnostic methods remain prohibitively expensive or inaccessible. This is particularly true in sub-Saharan Africa and South Asia, regions that bear the highest burden of TB.

Dr. Tamale believes this research demonstrates that academic institutions in Africa are not only capable of conducting high-impact scientific work, but also of contributing tangible solutions to the continent’s most urgent health challenges. “Victoria University is proud to be part of this breakthrough,” she said. “It proves that innovation isn’t about complexity—it’s about relevance, efficiency, and impact.”

The research also underscores Victoria University’s growing reputation as a regional leader in applied artificial intelligence and digital health. Under the guidance of scholars like Dr. Tamale, the university has positioned itself at the intersection of technology and public service, producing not only academic knowledge but also real-world applications.

The full paper, Improving Tuberculosis Detection in Chest X-Ray Images Through Transfer Learning and Deep Learning: Comparative Study of Convolutional Neural Network Architectures, is now accessible in JMIRx Med. A preprint was previously available on medRxiv, and the dataset used for training and evaluation remains open-source via Kaggle, inviting further research and replication.

For healthcare providers and policymakers seeking to modernize disease diagnostics without overhauling their technical infrastructure, the message is clear: simpler can be smarter. With VGG16 leading the way, the future of AI in global health may be both more accessible and more effective than previously imagined.