The experiments were conducted on a dataset made up of 2D axial slices extracted from BraTS2020. In this dataset, the 3D volumes were cut into 2D axial slices after extracting the brain mask and applying Z-score normalization. We only kept slices where the tumor proportion was greater than 2% of the brain volume for the positive class.

A slice is labeled as “tumor” when the proportion of segmented pixels (labels 1, 2, or 4) compared to the number of brain pixels in the same axial slice, defined by a non-zero FLAIR intensity, reaches 2%, and slices below this threshold are labeled “non-tumor.” Since BraTS2020 only includes patients with glioma, the “non-tumor” slices are therefore slices of tumor-containing brains where lesions are not visible.

As a result, the performances reported in this study reflect the model’s ability to locate slices with lesions within an exam already known to be tumorous. They do not allow any generalization regarding the ability to distinguish a tumorous patient from a healthy patient, nor any use for screening. Such an evaluation would require a dataset that includes control exams, which BraTS2020 does not provide; and this is a future direction for this work.

Table 1 shows the final distribution of the dataset after preparation, stratification, and class balancing. The train/validation/test split (70%/15%/15%) was done in a stratified way with a fixed random seed (SEED = 42) to ensure the experiments could be reproduced.

Table 1. Dataset distribution.

The seed value was set to SEED = 42; axial slices were extracted from the BraTS2020 volumes following non-zero Z-score normalization.

All images were resized to 224 × 224 pixels using bilinear interpolation. Pixel values were normalized to the [0, 1] range by dividing by 255. For each image, Z-score normalization was first applied to non-zero voxels (nonzero masking), followed by cropping to the brain region to eliminate the uninformative black background. The images were processed as single-channel (1-channel) grayscale images, consistent with our previous work reducing computational cost compared to RGB approaches.

LightBrainNet is based on the principle of depthwise separable convolutions, introduced by Howard et al. in MobileNet. Table 2 shows the neural network architecture, and you can see that the spatial resolution decreases from 224 pixels while the depth increases from 1 to 512. The GlobalAveragePooling2D is a key choice because if we had used a Flatten, we would have ended up with a very large vector of 14 × 14 × 512 = 100,352 values.

Table 2. Model architecture.

DW = DepthwiseConv2D (3 × 3 convolution applied channel-wise) · PW = Pointwise Conv2D (1 × 1 convolution for inter-channel mixing) · BN = BatchNormalization. Each DW/PW block follows the pattern DW → BN → ReLU → PW → BN → ReLU.

A comprehensive data augmentation strategy comprising nine techniques was applied exclusively during the training phase. Table 3 details each technique, its parameters, and its clinical rationale. This strategy builds upon findings from our previous work, which demonstrated that a combination of geometric and photometric techniques significantly improves the robustness of detection models on small MRI datasets.

Table 3. Data augmentation techniques.

All augmentations were applied on-the-fly during training only. No augmentation is applied to the validation and test sets.

Training was conducted in two phases, each with its own learning rate. The first phase (15 epochs, lr = 1e−3) was a warm-up phase during which the layers were trained starting from randomly initialized weights (Glorot uniform). The second phase (up to 30 additional epochs, lr = 1e−4) consisted of full fine-tuning with a reduced learning rate. We used Binary Cross-Entropy as the loss function, which is standard for binary classification. The Adam optimizer was employed with default parameters (β₁ = 0.9, β₂ = 0.999) alongside a ReduceLROnPlateau scheduler (factor = 0.5, patience = 7 epochs).

We employed three callbacks: ModelCheckpoint (saving the model that maximizes validation AUC-ROC), EarlyStopping (patience of 15 epochs, based on validation AUC-ROC), and CSVLogger (comprehensive logging for reproducibility). The batch size was set to 16 images. The random seed was fixed at 42 for TensorFlow, NumPy, and Python’s random module. Full training was conducted on an Intel i7 CPU (16 GB RAM, no GPU), in accordance with the targeted deployment constraint.

After the training phase, the weights remain fixed in FP32 (32-bit) format; therefore, post-training quantization is applied to convert the values into INT8 (8-bit) integers using TensorFlow Lite. This process aims to reduce the model size by a factor of four and accelerates data processing by the processor.

The final evaluation was conducted exclusively on the BraTS2020 test set (n = 5230 images). The decision threshold was optimized using the optimal Youden index (maximizing sensitivity + specificity − 1), calculated on the validation set. The reported metrics are:

overall accuracy, sensitivity, specificity, positive predictive value (Precision/PPV), F1-score, and AUC-ROC.

Table 4 presents the comparative performance achieved by LightBrainNet and the six baseline models on the independent test set, using the optimal Youden threshold (0.6395). LightBrainNet (FP32) achieves the best overall accuracy (96.85%), sensitivity (96.62%), specificity (97.17%), F1-score (97.30%), and AUC-ROC (99.51%) among all evaluated models, including ResNet50 and InceptionV3, while using 27× and 25× fewer parameters, respectively.

Table 4. Les performances comparatives des modèles.

Orange: models requiring a GPU for real-time inference. Green: models deployable on standard CPUs. The comparison is indicative and uncontrolled, as the results for the other models are drawn from the literature.

Between epochs 1 and 15, the validation curve fluctuates wildly, with loss rising to 0.9 and accuracy dropping to 0.67, while the training curve descends smoothly (See Figure 1). This contrast indicates not a data issue, but an excessively high learning rate (lr = 10⁻³). From epoch 15 to 45, the learning rate was reduced (lr = 10⁻⁴); the validation curve then converged with the training curve and ceased to fluctuate. No overfitting was observed, demonstrating that the model’s architectural lightness did not come at the cost of capacity. With training accuracy matching validation accuracy (96.85%), the training-validation gap is 0.00%, further confirming the absence of overfitting, unlike the gaps observed for the sequential CNN (8.34%) and ResNet50 (4.47%) in our previous work. The final ROC curve shows an AUC of 0.9951 with an optimal Youden threshold of 0.6395, corresponding to a sensitivity of 96.62% and a specificity of 97.17%. This sensitivity level, which enables the classification of 96.62% of true tumors, is clinically satisfactory for a first-line screening tool designed to prioritize examinations requiring expert review.

Figure 1. Training curves.

Table 5 presents an analysis of the compatibility between LightBrainNet INT8 requirements and the typical characteristics of computing equipment available in district hospitals and university teaching hospitals (CHUs) in sub-Saharan Africa, based on data from the Survey on Hospital Technological Resources in West Africa (ERTHA, 2023).

Table 5. Compatibility of LightBrainNet INT8 with African hospital equipment.

Hardware requirements based on actual inference measurements (Intel i7, 16 GB RAM). Hospital profile based on the 2023 ERTHA survey (n = 47 facilities, 8 sub-Saharan African countries).

The analysis demonstrates full compatibility between LightBrainNet INT8 and the hospital equipment typically found in the sub-region. On a computer equipped with an i7 processor and 16 GB of RAM, processing a complete scan (30 slices) takes just 0.4 seconds, a clinically negligible duration that integrates seamlessly into the radiology workflow. Its 1.04 MB size allows the model to be distributed via a standard USB drive and deployed offline, eliminating any reliance on a stable internet connection.

The LightBrainNet tool was designed to prioritize brain imaging examinations that have already been ordered, rather than as a mass screening device or a substitute for radiologist expertise. The proposed workflow consists of four steps:

First, acquisition of the brain MRI, including a FLAIR sequence, the only modality on which our model was trained and validated. Second, automatic extraction and preprocessing of axial slices from DICOM files, involving Z-score normalization of brain voxels and resizing to 224 × 224 pixels to replicate training conditions. Third, automatic analysis by LightBrainNet, taking 0.4 seconds for a 30-slice examination on a standard processor without a GPU, followed by the display of the probability of tumor presence per slice and the flagging of slices exceeding the decision threshold. Fourth, prioritization of flagged examinations for review by a specialist physician.

To the best of our knowledge, LightBrainNet is the first deep learning-based brain tumor classification pipeline specifically designed, trained, and validated for deployment on CPUs, without the need for GPUs, including an explicit analysis of clinical compatibility in resource-constrained environments. This contribution directly addresses the gap identified in our systematic review, which surveyed 89 studies from 2019 to 2024 without finding any that resolved this specific challenge.

The performance of LightBrainNet (AUC-ROC = 99.51%) surpasses that of ResNet50 (98.81%), a result obtained in our previous work under comparable conditions. This superiority, seemingly surprising for a model with 26 times fewer parameters, can be attributed to several factors. First, the depthwise separable architecture enforces a factorization of representations that acts as implicit regularization, thereby reducing overfitting. Second, the more comprehensive data augmentation strategy (9 techniques versus 5 in the previous study) enhances robustness.

Thirdly, GlobalAveragePooling2D, being more resistant to overfitting than Flatten + Dense, contributes to better generalization.

The consistent progression across our four articles, from the study of augmentation and architectural comparison to the systematic literature review and the proposal of an original, deployable solution, constitutes a rigorous and coherent research trajectory, grounded in the real-world needs of healthcare systems in sub-Saharan Africa.

Although no study explicitly targets the same objective as LightBrainNet, indirect comparisons can be made with work utilizing MobileNetV2 or EfficientNet-B0 for medical image classification. MobileNetV2, applied to MRI classification in our comparative study, achieves 94.76% accuracy, 2.09 percentage points lower than LightBrainNet. The primary architectural difference lies in MobileNetV2’s use of inverted residual blocks; while effective, these were developed for natural RGB images and lack the specific optimizations suited to single-channel grayscale medical images. Our architecture, specifically designed for this context, beats MobileNetV2 by 2.09 points while having fewer parameters (1.04MB vs 3.5MB).

Several limitations of this study should be acknowledged:

First, the binary classification (tumor/non-tumor) does not distinguish between histological types (meningioma, glioma, and metastasis), a limitation inherent to the chosen task, yet justified by the objective of initial detection in patients who have undergone brain MRI.Second, performance was evaluated using data from the same training distribution (the BraTS2020 source); this means that external validation using MRI data from African hospitals remains essential before any actual clinical deployment.Third, training on a CPU (Intel i7, without a GPU) limited the number of experiments that could be explored in this study; training on a GPU could potentially improve performance by allowing for more exhaustive hyperparameter tuning.Fourthly, decision explainability (Grad-CAM, SHAP) has not been integrated into the current version of the pipeline; this is a significant limitation for clinical acceptance.Fifthly, the BraTS2020 volumes used were provided already co-registered and resampled. Clinical deployment based on raw DICOM data would require a complete preprocessing pipeline, the impact of which on performance has not been evaluated.

Several extensions to this work are envisaged, namely:

Integrating Grad-CAM activation maps [4] into the inference pipeline to visualize the MRI regions that led to the decision; this would increase clinician confidence.Extending the system to multi-class classification (glioma/meningioma, etc.) is feasible by adapting the output layer.Prospective validation using data collected at partner hospitals in West Africa represents the next critical step in validating the system’s actual clinical relevance.