The remainder of this paper is structured as follows: Section 2 sheds light on related work, Section 4 presents the methodology, including data preprocessing, model architecture, and training strategy. Section 5 describes the experimental setup and hyperparameters. Section 6 provides the results and detailed discussion of the model's performance. Finally, Section 7 concludes the paper and discusses future research directions.

The overall structure of the ESA-YOLOv5m model is shown in Figure 2. It builds upon the YOLOv5m backbone, with the ESA module placed immediately after the Spatial Pyramid Pooling-Fast (SPPF) layer. This placement enhances the spatial discrimination capability of the network by amplifying tumor-relevant regions and suppressing non-informative background details. The model pipeline consists of four functional stages: feature extraction, spatial refinement, multi-scale feature fusion, and prediction. This configuration enables the model to efficiently detect tumors of varying sizes and contrast levels while maintaining real-time performance.

The model receives T1-weighted contrast-enhanced MRI slices as input. Preprocessing involves intensity normalization, resizing to a fixed input resolution, and data augmentation operations such as flipping, rotation, HSV shifting, mosaic, and mixup. These augmentations improve generalization and robustness without distorting anatomical structures. Each image is resized to 832 × 832 pixels to balance fine-detail preservation and memory efficiency.

The structure's backbones are built off of the hierarchical, low- and high-level feature maps of the CSP-Darknet53 used in YOLOv5m. The Enhanced Spatial Attention module is added directly after the SPPF, which is where the features are pooled and where this integration refinement takes place. The refinement focuses on the activation maps of the highly background-attenuated regions by boosting the tumor-specific activations. The attention mechanism improves the feature representation and, at the same time, improves the network's sensitivity to smaller and low-contrast tumor features with marginal added computation. The traditional attention designs, for example, Squeeze-and-Excitation (SE) (28) block and the Convolutional Block Attention Module (CBAM) (29) improve the feature representation with channel or sequential channel-spatial processing. Nevertheless, SE considers only the channel recalibration and disregards spatial relevance, whereas CBAM uses attention in order to result in more computations. Conversely, the suggested ESA introduces adaptive spatial weighting in a non-iterative manner by multi-scale convolutions, which enables the network to maintain local and global dependencies without many extra parameters. The design results in the localization of tumor boundaries more sharply, and the interpretation is better within the restrictions of real-time. In Table 2, the proposed ESA is distinguished by conventional attention modules at multiple scales by considering only spatial relationships with multi-scale convolutional fusion, which can localize successfully with limited computational cost.

Following the ESA-enhanced backbone, the neck uses a Cross Stage Partial Path Aggregation Network (CSP-PAN) to merge features from multiple scales. This ensures that both global context and localized features are retained, allowing accurate detection of tumors regardless of their size or location. The detection head, identical to YOLOv5m's standard head, outputs bounding boxes, class probabilities, and confidence scores for the three tumor categories: glioma, meningioma, and pituitary. The ESA-refined features contribute to tighter bounding boxes and improved confidence in tumor localization.

The model is trained end-to-end using a composite loss function that combines classification, objectness, and localization terms. The total loss is defined in Equation 1:

In Equation 1, L_cls represents the classification loss, L_obj is the objectness loss, and L_CIoU denotes the Complete IoU loss. The weighting coefficients λ_cls, λ_obj, and λ_loc balance the contributions of these components. The inclusion of the Complete IoU term improves the geometric alignment between predicted and ground truth bounding boxes, leading to more accurate tumor localization.

The entire network is optimized using the loss function described in Equation 1. Training employs an AdamW optimizer with cosine learning rate scheduling and standard YOLOv5 augmentation strategies. This configuration ensures efficient convergence and balanced optimization across all detection components. By integrating the ESA module into the YOLOv5m backbone, the proposed model achieves refined spatial attention with negligible computational overhead. The attention mechanism enhances the focus on diagnostically significant tumor areas, thereby improving detection precision and recall. The improved feature representation also reduces false positives and accelerates training convergence. Overall, the ESA-YOLOv5m framework maintains the lightweight and real-time characteristics of YOLOv5m while delivering substantial gains in spatial accuracy and interpretability, making it suitable for deployment in clinical diagnostic systems.

Let the MRI dataset be defined as D={I1,I2,…,In}, where each image I_i contains annotated tumor regions. Preprocessing transforms each raw image through normalization, resizing, and augmentation operations, collectively represented by T(·) in Equation 2.

Here, N denotes intensity normalization, R represents resizing to 640 × 640 pixels, and A applies augmentations such as flipping, rotation, HSV jittering, mosaic, and mixup. This transformation standardizes pixel intensity and improves generalization across subjects and MRI scanners. The YOLOv5m backbone extracts features F_t for each batch at epoch t. The ESA module refines these features according to Equation 3:

In Equation 3, f₁, f₂, and f₃ denote convolutional layers, δ(·) is the SiLU activation, and σ(·) the sigmoid function generating the spatial weighting map. The enhanced map Ft′ strengthens tumor-relevant activations and suppresses background noise before detection. Predictions consist of bounding boxes B^, class probabilities Ĉ, and confidence scores Ŝ, obtained as Equation 4.

The model minimizes a composite loss given by Equation 5:

Parameters are updated iteratively using the AdamW optimizer using Equation 6.

The training procedure is expressed compactly in Equation 7:

The proposed framework uses the Figshare Brain Tumor MRI dataset, containing 3,064 T1-weighted contrast-enhanced MRI images from 233 subjects. The dataset includes three tumor types—glioma, meningioma, and pituitary—with resolutions of 512 × 512 pixels. Each image is resized to 640 × 640 during training and augmented using transformations defined in Equation 2. Dataset statistics are summarized in Table 3.

As part of the preprocessing stage, binary tumor masks assist in confirming the spatial correspondence between the tumor and the respective annotated bounding box. These are the examples presented in Figure 3. The top row contains raw MRI slices, whereas the bottom row contains the respective masks. These masks serve the purpose of assuring the proper placement of bounding boxes around tumor margins before the training stage.

Upon completion of the verification process, the labeled MRI images are prepared for YOLOv5 training, as presented in Figure 4. Each of the red bounding boxes represents boxes of detected tumorous areas, and the numbers 1–3 refer to the classes of glioma, meningioma, and pituitary. This visual representation proves the persistent annotation agreement in the entire dataset.

The YOLOv5m backbone employs a CSP-Darknet53 structure for multi-scale feature extraction. The ESA module is integrated immediately after the SPPF layer, enhancing spatial feature discrimination. The ESA operation can be mathematically expressed as:

In Equation 8, ⊙ denotes element-wise multiplication, C represents a convolutional mapping, σ is a sigmoid activation producing spatial attention weights, and α is a learnable scaling coefficient. This mechanism emphasizes high-importance tumor regions and suppresses irrelevant background activations. Because ESA is applied only once per forward pass, its computational overhead remains minimal. Algorithm 1 describes the stage of refining the feature of the ESA block prior to inserting it into the YOLOv5m backbone.

Transfer learning is applied using YOLOv5m weights pretrained on the COCO dataset. The model is fine-tuned for 150 epochs using the AdamW optimizer with cosine learning rate scheduling. The dynamic learning rate η_t varies as per Equation 9:

Here, η₀ is the initial learning rate, t is the current epoch, and E is the total number of epochs. The functionality of reducing learning rates is to eliminate uncertainty and stabilize training. The batch size is configured to be between 8 and 16. All experiments are repeated with multiple random seeds to ensure reproducibility.

Model performance is quantified using Precision (P), Recall (R), and mean Average Precision at IoU thresholds 0.5 and 0.5:0.95. For each tumor class c, the average precision AP_c is computed over the precision-recall curve, and mean mAP is given by Equation 10:

Where C is the total number of classes, inference latency (milliseconds per image) is also measured to assess real-time feasibility.

The incorporation of ESA into the backbone model of YOLOv5m increases the total parameters to just under 4%. ESA retains latency under ten milliseconds per image, signifying real-time performance, and maintains exemplary results with little additional computation. The added ESA has been shown to effectively enhance spatial attention, yielding better Precision, Recall, and mAP values. These traits, along with the framework's features, make the model especially advantageous for use in clinical decision-support systems and for deployment on edge devices. The proposed approach has developed an explainable and sophisticated detection pipeline encompassing efficient data preprocessing, lightweight attention modules, and robust evaluation metrics. In Figures 3, 4, the qualitative examples demonstrate the accurate and efficient tumor detection with few false positives, achieved via the model, due to the extensive dataset and annotation pipeline.

The experimental corpus was derived from the Figshare Brain Tumor MRI dataset, which comprises three classes: glioma, meningioma, and pituitary. Using stratified sampling to maintain the distribution of the classes, the dataset was split into training (80%) and validation (20%) subsets. Every image was resized to 640 × 640 pixels before being sent to the data loader. Training and validation data were processed with the same normalization pipeline, while data augmentation was applied only to the training set. The training was performed for 150 epochs with an initial learning rate η₀ = 0.001 following a cosine annealing schedule defined in Equation 11. The AdamW optimizer was used with β₁ = 0.9, β₂ = 0.999, and a weight decay of 1 × 10⁻⁴. A batch size of 8 was used for T4 GPUs and 16 for the A100 configuration. Gradient accumulation was enabled to stabilize optimization during small-batch training, and mixed-precision computation (FP16) was used to accelerate convergence while minimizing GPU memory consumption. Each epoch computed detection losses (classification, objectness, and CIoU) as per Equation 1. The best-performing model was selected based on the highest mean Average Precision at 0.5 IoU (mAP@0.5) on the validation set.

Preprocessing involved intensity normalization and resizing to maintain a uniform image scale. Augmentation techniques were applied to increase the diversity of training samples and improve generalization to unseen MRI scans. These transformations included random horizontal flips with probability p = 0.5, HSV hue-saturation-value jitter (±0.015), random rotation (±10°), Mosaic augmentation (probability p = 0.5), and Mixup blending (α = 0.2). Each augmented sample was normalized to the [0,1] range and standardized using mean subtraction and variance scaling. The final preprocessed tensors were formatted into batches of shape (B, 3, 640, 640), where B is the batch size. Each label file was encoded in YOLO format {x_c, y_c, w, h, c}, corresponding to the bounding box center coordinates, width, height, and class ID.

Hyperparameter tuning was performed empirically across multiple runs. The learning rate η followed the cosine annealing schedule, using Equation 11.

Where E is the total number of epochs and t denotes the current epoch index. Momentum was set to 0.937, and the weight decay coefficient was 1 × 10⁻⁴. The object confidence threshold for detection was fixed at 0.25, and the Non-Maximum Suppression (NMS) IoU threshold was set to 0.45. During training, a label smoothing factor of 0.1 was used to prevent overfitting and enhance model calibration. The overall loss function L_total in Equation 1 was optimized using backpropagation with automatic mixed precision (AMP) to accelerate computation. Early stopping was employed with a patience of 30 epochs, halting training when no improvement in mAP@0.5 was observed.

Evaluation metrics included Precision (P), Recall (R), mean Average Precision at IoU thresholds 0.5 (mAP@0.5), and averaged between 0.5:0.95 (mAP@0.5 0.95). Inference speed (milliseconds per image) and GPU utilization were recorded to assess computational efficiency. For each trained model, performance was evaluated using five random seeds to confirm reproducibility. Statistical mean and standard deviation values were reported across these runs. Visualization of predictions was performed using Matplotlib, where each detected bounding box was annotated with tumor class, confidence score, and bounding box coordinates. Confusion matrices and Precision-Recall curves were generated to analyze inter-class performance.

Reproducibility, consistency, and equity in assessment are maintained by the thorough setup described in Table 4. The variability in tumor forms and levels of brightness in the Figshare dataset has become a standard for the assessment of medical detection models. The mixed-precision training, AdamW optimizer, and cosine annealing postpone convergence and eliminate the risks of gradient vanishing. The ESA module improves detection performance at practically no extra processing expense, which supports real-time inferences on cloud and edge devices.

Figure 5 presents a comparison of the performance metrics of the baseline and ESA-enhanced YOLOv5m models. The baseline model achieved an mAP@0.5 of approximately 0.80 for the mAP, Precision, and Recall metrics, while the ESA-enhanced model exceeded 0.90 for each metric, indicating a large increase in detection capacity. This feature of the model is likely due to the Enhanced Spatial Attention (ESA) module, which concentrates on accentuated tumor areas and background suppression, which improves the accuracy and sensitivity of the detection. Furthermore, the increase in mAP@0.5 indicates an improvement in the model's ability to detect and accurately assess the number of tumors present, while Precision and Recall also demonstrate an increase in the number of falsely detected and missed tumors. The detected improvement of 11%–12% mAP on the clinical side translates to about 1–2 missed tumors per 100 cases compared to the baseline detector per each clinical mAP improvement. More confident and rapid triage decisions, particularly on smaller or early lesions that are often missed during manual assessments, translate to improved diagnostic accuracy.

Figure 6 shows both models' mAP@0.5 on training epochs. The baseline model takes a while to converge. After 50 epochs, it hovers around 0.80 mAP, while the ESA model is on a much steeper learning curve, surpassing 0.85 mAP by epoch 15 and stabilizing at 0.90 mAP by epoch 50. The ESA model performed both faster and better, indicating that ESA does improve feature representation, especially for small, complex tumor boundaries. The ESA block incurs minimal computational cost and is therefore advantageous in medical imaging, as it focuses on the more relevant tumor structures and delivers significant performance improvements. Fine-grained structure localization accuracy is of utmost importance.

The classification behavior of models is further explained by the Precision and Recall indicators. We observe the training progresses for Precision and Recall of the baseline and ESA models in Figure 7. The baseline models show close to 0.80 for both metrics, while the ESA models are above 0.90 and consistently outperform the baseline models. The ESA models exhibit fewer false positives and, hence, higher precision, meaning they are more confident in identifying and accurately delineating regions containing tumors. The attention mechanism effectively suppresses background noise and emphasizes the reconciling boundaries of tumors. The ESA models also show higher recall, meaning they are more capable of identifying and retrieving a majority of the actual tumors with the least possible misses. This is critical in most medical cases where a failure to recognize a tumor can cause serious problems with the diagnosis.

Table 5 captures the results from the last phase of each model's performance metrics. The ESA-enhanced model captures an 11%–12% elevation in overall detection performance. This enhancement is sustained through all the runs and random seeds, which confirms the reliability and acceptability of the approach.

Qualitative visualizations further confirm the effectiveness of ESA. The ESA-enhanced model produces more accurate bounding boxes, particularly for small or irregularly shaped tumors. It also exhibits stronger confidence scores and fewer background misclassifications. The baseline model tends to under-detect low-contrast tumors or merge multiple lesions into a single bounding box, whereas the ESA model resolves these ambiguities effectively.

ESA's integration adds only a lightweight spatial attention mechanism, with no marked increase in model size nor inference duration. The average latency in image processing is below 10ms, and hence, retains its applicability and speed in real-time clinical usage on an NVIDIA A100 GPU. This clearly shows improvement in performance with no loss in efficiency.

In order to enhance the verification process of the ESA-YOLOv5m architecture, additional controlled experiments were designed and executed under the same primary training setup parameters. This encompasses ESA placement ablation, class-wise analysis, cross-validation stability, and assessment of the efficiency and stability of ESA placement. The additional results retain the coherence with the primary results as discussed in Section 6.

Grad-CAM visualizations were produced on sample MRI slices to estimate the attention to clinically relevant areas of the proposed ESA-YOLOv5m model. The heatmaps in Figure 12 show that ESA-YOLOv5m exhibits a more localized and concentrated activation in the tumor core and its pathological areas than the baseline YOLOv5m model. This increased transparency gives a visual representation of how the model makes decisions, and thus, enhances trust and helps to integrate the model into diagnostic processes.

The inclusion of the ESA layer immediately after the SPPF block of the ESA-enhanced YOLOv5m model increases feature map responsiveness by allowing the model to adaptively focus on critical tumor areas of the feature maps, attenuating the less important background. This contrast between the ESA-enhanced YOLOv5m model and the baseline YOLOv5m snapshot.

These results demonstrate that the proposed model provides measurable performance gains over prior CNN and YOLO-based approaches while preserving computational efficiency. As summarized in Table 10, the ESA-YOLOv5m model surpasses the best-performing prior attention-based models across multiple key metrics with minimal overhead, making it more suitable for real-time clinical applications. In contrast to prior works that often prioritized accuracy at the expense of speed and practicality, the proposed ESA-YOLOv5m achieves both high accuracy and real-time performance. Its lightweight attention mechanism offers a better balance between model complexity and clinical applicability. This makes it a promising candidate for deployment in real-world diagnostic workflows, enabling earlier detection and improved treatment outcomes for patients with brain tumors.

The experimental results confirm that incorporating ESA into the YOLOv5m architecture significantly improves brain tumor detection performance. The enhanced mAP, Precision, and Recall highlight the network's improved ability to focus on diagnostically important features while minimizing false alarms. This is especially valuable for healthcare scenarios, where accurate and fast tumor detection can assist radiologists and reduce manual screening time. The ESA model, in addition to demonstrating stable convergence, shows consistent improvements across runs, which is a promising sign for real-time deployment within medical imaging workflows. The ESA model's convergence also enables it to be scaled for edge deployments or fused within hospital ecosystem diagnostic lite systems. These results confirm that lightweight attention modules, when applied in a judiciously designed framework, can considerably enhance object detection in highly specialized tasks, such as brain tumor imaging, with maintained speed and scalability.

Even though the results are promising, several limitations with the dataset must be addressed. The Figshare dataset presents class imbalance (i.e., glioma > meningioma), scanner-specific acquisition parameters, and poor demographic variety, all of which may affect generalizability across other institutions. Future work should focus on validating ESA-YOLOv5m on multi-center, multi-site, and multi-vendor cohorts to address these biases. Furthermore, the proposed ESA-YOLOv5m framework is currently set up to work on separate 2D MRI slices, although it is capable of high detection results. This method lacks the ability to take volumetric continuity between adjacent slices, which may be of significance to identifying diffuse or infiltrative tumor structures. In the future, we will extend ESA-YOLOv5m to 3D volumetric tumor detection and multi-modal fusion (MRI, CT, and PET), such that holistic spatial-temporal tumor characterization can be used to achieve more precise diagnostic results.