TITLE:
LightBrainNet: A Lightweight Deep Learning Model for Identifying Tumor-Bearing MRI Slices in Resource-Limited Settings
AUTHORS:
Gedeon Nitiema, James T. Kouawa, Ahmed Dooguy Kora
KEYWORDS:
Brain Tumor Classification, MRI, Lightweight Deep Learning, Depthwise Separable Convolution, INT8 Quantization, CPU Deployment, Resource-Constrained Environments, Sub-Saharan Africa
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.11,
November
10,
2026
ABSTRACT: The detection of brain tumors using deep learning has reached performance levels close to expert radiologists on standardized benchmarks. However, state-of-the-art models like ResNet50 and VGG16 require expensive GPUs, which are inaccessible in most hospitals in sub-Saharan Africa. To our knowledge, no study has so far proposed a classification pipeline deployable on a standard CPU. We present LightBrainNet, a lightweight architecture based on depthwise convolutions, designed for binary classification of grayscale brain MRI images. The model (around 1 million parameters, 3.65 MB) was trained on 34,854 axial slices and optimized with INT8 quantization, achieving a CPU latency of 12.1 ms per image. LightBrainNet achieves an accuracy of 96.85%, a sensitivity of 96.62%, a specificity of 97.17%, an F1 score of 97.30%, and an AUC-ROC of 99.51%, outperforming ResNet50 by +1.85% in accuracy with 27 times fewer parameters. The INT8 version reaches 96.85% accuracy with just 16 GB of RAM. LightBrainNet provides a precise, lightweight solution and has been deployed in a GPU-free environment for brain tumor screening in resource-limited settings. The detection results are promising, which points to its potential use in hospitals in sub-Saharan Africa.