Deep Learning-Based Brain Tumour Detection and Classification

Early and accurate detection of brain tumours is essential for improving patient outcomes and guiding effective treatment plans. Traditional diagnostic procedures such as biopsies, performed after brain surgery, are invasive and carry significant risk. With recent advancements in technology, deep learning techniques now offer a non- invasive alternative capable of supporting radiologists in interpreting MRI scans more efficiently and consistently. It is important to emphasize that AI-based brain tumour detection systems are not intended to replace radiologists or act as autonomous diagnostic tools. Rather, they function as clinical decision-support systems designed to enhance, not substitute, expert judgement. Numerous studies highlight that deep learning models can reliably pre-screen large volumes of MRI scans, flag abnormal regions and prioritize high-risk cases, thereby reducing radiologist workload and improving diagnostic consistency. In a typical clinical workflow, where a radiologist may manually review approximately 20 MRI scans per day, AI-assisted triaging enables the same clinician to process significantly larger volumes, often 3–5 times more since the model performs initial detection while the radiologist focuses on verification and final interpretation. Thus, the proposed CNN model serves as an efficient double-verification system that assists clinicians in identifying subtle tumour signatures, reduces oversight risk, and improves overall diagnostic throughput while maintaining human oversight at every stage. In this study, we develop a deep learning–based approach using a customized Convolutional Neural Network (CNN) for the classification of brain tumours into 4 classes - Meningioma, Glioma, Pituitary tumour and no tumour. The model was trained on a dataset of 5,788 MRI images and achieved an accuracy of 98.48%, outperforming several existing models trained on the same dataset as reported in the literature, including VGG-16 (97%), CNN-SVM (95.16%), IVX-16 (96.94%), XGBOOST (90%), an 8-layer CNN (96.86%), and a 3D CNN model (98.03%). The proposed model demonstrates strong capability in identifying tumour-specific visual patterns and provides rapid, reliable predictions and using XAI technique like GradCAM+ offering valuable decision- support to clinicians. By reducing dependency on manual interpretation alone, this work highlights the potential for deep learning methods to enhance diagnostic accuracy, reduce variability between practitioners, and facilitate earlier intervention. Overall, the study contributes toward the meaningful integration of artificial intelligence into routine medical practice, making diagnostic processes both smarter and more accessible.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Liza, Ansu

Advisor dc:contributor.advisor

• Coetzer, Johannes

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136284

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136284