In this study, we systematically developed and evaluated multiple lightweight and efficient deep learning models specifically, ResNet-18 (both pretrained on ImageNet and trained from scratch), ResNet-34, ResNet-50, and a custom CNN for detecting glioma, meningioma, pituitary tumors, and non-tumorous cases using conventional MRI images. By conducting a side-by-side comparison, our work offers a granular view of how different levels of network depth and parameter counts affect diagnostic accuracy, especially when computational resources are limited. This focus on moderate-depth, high-efficiency architectures directly addresses the needs of smaller clinical centers where GPU power may be scarce yet accuracy demands remain high. Our research demonstrated that lightweight models, such as ResNet-18, can achieve exceptionally high performance through transfer learning equalling or in some cases surpassing the results of more parameter-heavy or complex architectures typically found in state-of-the-art (SOTA) models.

Recent deep learning models for brain tumor classification have demonstrated remarkable accuracy, often exceeding 98%, by incorporating advanced architectural enhancements such as attention mechanisms, generative augmentation, and hybrid feature extraction techniques. However, these enhancements often come at the cost of increased computational complexity, making them less suitable for clinical environments with limited resources. For instance, SAlexNet [12] integrates a Hybrid Attention Mechanism (HAM) and residual layers to enhance feature extraction, achieving an accuracy of 99.69%. HAM helps SAlexNet capture both spatial and channel-wise dependencies, allowing it to enhance tumor localization and segmentation. Similarly, NeuroNet19 [10] builds upon VGG19 with an Inverted Pyramid Pooling Module (iPPM) to capture multi-scale features, attaining 99.3% accuracy. TumorGANet [9] employs ResNet-50 in conjunction with a GAN-based data augmentation strategy, reporting a classification accuracy of 99.53%. Additionally, Qureshi et al. [11] introduced an ultra-light CNN that combines deep learning with Gray-Level Co-occurrence Matrix (GLCM) texture analysis, reaching an accuracy of 99.23% while maintaining efficiency. While these models achieve high performance, their computational demands vary. SAlexNet and NeuroNet19 incorporate attention mechanisms and pooling modules, which may increase inference time, whereas TumorGANet relies on a GAN-based augmentation strategy that can add training complexity. However, Qureshi et al.'s ultra-light CNN emphasizes efficiency, demonstrating that smaller architectures can still achieve high accuracy. In contrast, our study focuses on "off-the-shelf" ResNet architectures, specifically ResNet-18, which balances simplicity with performance. Our approach demonstrates that strong data augmentation strategies, such as random affine transformations, color jittering, and rotation augmentation, can achieve SOTA performance without extensive architectural modifications, making it more practical for real-world deployment in smaller clinical settings. Moreover, when compared to traditional machine learning approaches such as artificial neural networks (ANNs) and random forests, our ResNet models demonstrates a clear advantage. For example, Hamd et al. [19] reported an AUC of 0.97 using ANN for pediatric brain tumor detection. In contrast, our ResNet-18 model achieved an AUC of 1.0 for glioma classification, underscoring the benefits of convolutional neural networks in learning spatial patterns and textures directly from MRI scans. The residual connections in ResNet further facilitate stable training by mitigating the vanishing gradient problem, enhancing feature extraction across deeper layers. Another key differentiator of our study is the systematic evaluation of multiple ResNet architectures, including ResNet-18, ResNet-34, and ResNet-50, to assess the impact of network depth and transfer learning on classification performance, providing a comparative perspective on computational efficiency versus accuracy. Unlike SAlexNet or NeuroNet19, which introduce additional attention and pooling mechanisms, our pipeline maintains a standardized architecture while leveraging strong data-level augmentations. This architectural simplicity enables easy integration into existing hospital frameworks, as ResNets are widely supported across various deep learning platforms used in clinical AI applications [20].

The ResNet model, a deep residual network architecture developed by He et al. [21], was pre-trained on the ImageNet dataset comprising over 14 million images. The use of ResNet allowed us to exploit its advanced feature extraction capabilities, resulting in more accurate and reliable detection of brain tumors compared to traditional ANN approaches. The higher AUC achieved by our CNN-based approach underscores the importance of using specialized deep-learning models like CNN with utilizing pre trained model such as ResNet-18 for medical image analysis. These models, pre-trained on extensive datasets, provide a robust framework for tackling complex medical imaging challenges, ultimately leading to improved diagnostic accuracy and better patient outcomes.

The reason we chose the CNN model for our model, was due to the CNN architecture is specifically designed to handle image data by leveraging layers that perform convolution operations to detect features such as edges, textures, and shapes within the images. These convolutional layers scan the MRI images with multiple filters to create feature maps that highlight important patterns. Pooling layers then reduce the dimensions of these feature maps, retaining critical information while minimizing computational load. Finally, fully connected layers combine these features to classify the images into their respective categories. This hierarchical feature extraction process allows CNNs to effectively capture the spatial and structural nuances of MRI scans, leading to accurate detection and differentiation of various brain tumor types. The ability of CNNs to automatically learn and extract relevant features from complex medical images makes them particularly suitable for medical image analysis, resulting in high diagnostic accuracy and reliability.

The application of AI in neuro-oncology has proven not only to enhance diagnostic accuracy but also to significantly improve efficiency. A notable competition involved a brain tumor diagnosis comparing the performances of human radiologists and an AI system developed by the Artificial Intelligence Research Centre for Neurological Disorders and Capital Medical University. In the competition the AI system, Biomind, achieved an 87% accuracy rate, diagnosing 195 out of 225 cases correctly within 15 min. In contrast, a team of 15 radiologists manually diagnosed 148 cases correctly, achieving a 66% accuracy rate over 30 min [22]. This demonstrates AI's potential to revolutionize diagnostic processes in neuro-oncology by delivering faster and more accurate results.

Based on the confusion matrix, we observed that 12 MRI images of glioma were incorrectly predicted as meningioma. Patel et al. reported similar diagnostic challenges, describing two cases where MRI initially suggested meningiomas, but surgical outcomes confirmed glioblastomas [23]. They attributed these diagnostic errors to gliomas exhibiting MRI characteristics typically associated with meningiomas, such as the dural tail sign, CSF cleft sign, and broad dural contact. These imaging mimicry patterns could account for the misdiagnoses observed, suggesting that even with the application of AI, misdiagnosis is possible due to the inherent similarities between diseases. This highlights a limitation within AI applications in neuro-oncology that must be addressed in future studies. Additionally, it's crucial to acknowledge that medical images often face issues like artifacts and resolution degradation, which can impact the accuracy of machine learning-based diagnostic methods.

The utilization of AI in neuro-oncology holds transformative potential by transitioning from exclusive reliance on radiologist expertise to standardized and efficient diagnostics accessible worldwide. This is facilitated by cloud-based platforms that provide virtual machines, making the primary requirement for utilizing the model just a basic computer. Wahl et al. highlight how expert systems like our deep learning model could enhance healthcare in resource-limited settings [24]. Such systems can assist physicians with diagnostics and treatment decisions, similar to practices in wealthier nations, and even substitute for human experts in areas where they are scarce [23]. This is particularly valuable in under-resourced communities. However, a significant challenge remains in these settings—the limited availability of MRI machines [25]. Addressing this issue requires a dual approach: integrating existing AI technologies to mitigate the absence of medical experts and increasing the accessibility of MRI technology in low-income countries.

Our study's strengths include the use of a pre-trained ResNet-18 CNN, which provided high diagnostic accuracy and AUC in classifying glioma, meningioma, pituitary tumors, and non-tumor MRI images. Utilizing publicly available Kaggle datasets ensures replicability and transparency. However, limitations exist, such as the use of 2D MRI slices, and potentially missing critical volumetric data from 3D images. The dataset's lack of clinical diversity may affect generalizability. Furthermore, the lack of longitudinal data and the necessity for additional clinical validation on larger databases underscore areas requiring future research.