Reliable and explainable classification of brain tumors via magnetic resonance imaging (MRI) is essential for early detection, prognostication, and tailored therapy in neuro-oncology. Diagnosis via human assessment is time-consuming, subject to inter-reader variation, and can be particularly challenging when tumor boundaries, contrasts, or morphology are irregular. MRI offers rich soft tissue contrast and non-invasive imaging, but image heterogeneity, noise, variation in acquisition parameters, and limited annotated datasets constrain model performance [1, 2]. Deep supervised learning methods (convolutional neural networks, transfer learning) have shown high accuracy in brain tumor MRI classification in recent years. Variants of EfficientNet (e.g. B0, B1, B2) have been used as backbones with success [3, 4, 5]. For example, "Multi-class classification of brain tumor types from MR images using EfficientNets" proved that fine-tuning EfficientNet variants (B0-B4), with the help of Grad-CAM visualization, achieved test accuracies ~ 98.8% and reasonable precision/recall across classes [6]. But still, supervised models have difficulty with the requirements of large and balanced labeled sets, overfitting in small sample regimes, and poor interpretability of learned features [7]. Self-supervised learning (SSL) presents a way to alleviate data annotation bottlenecks by learning powerful representations from unlabeled MRI slices. Contrastive SSL methods like SimCLR enable networks to become robust to variations and learn invariant features without explicit labels [8]. When SSL is combined with careful augmentation, feature normalization, and strong backbones, it can yield improved generalization in medical imaging tasks where labels are expensive or sparse. Interpretability is also crucial: models used in medical settings must allow clinicians to verify that predictions derive from medically meaningful image regions. Methods like Grad-CAM help visualize what regions the network is focusing on, increasing trust and facilitating error analysis [5, 9]. Without interpretability, even high-accuracy models may not be clinically accepted. In this study, we propose a framework combining SSL using SimCLR with EfficientNetB3 backbone, followed by supervised fine-tuning, to classify four classes of brain tumor plus no tumor. We also integrate Grad-CAM to ensure that decisions are spatially understandable. Our experiments show high accuracy (~ 98.3%), strong per-class metrics even on challenging tumor types, and interpretability that aligns with radiological features.

The major contributions of this work include:

The rest of the paper is organized as below. Section 2 provides an overview of related work in SSL, MRI tumor classification, and interpretability techniques. Section 3 outlines datasets, preprocessing, model architecture, and training. Section 4 outlines experiments, results, interpretability analysis, limitations, and future works. Section 5 concludes with a summary of findings and implications for clinical application.

Altin Karagoz et al. [10] introduced a self-supervised learning hybrid model called ResViT for the classification of brain tumors. The model integrates CNNs for local feature extraction with Vision Transformers for learning global contextualized representations. They also introduced synthetic MRI synthesis as a pretext task to class-wise balance distributions prior to fine-tuning for downstream classification. It was tested and trained on Figshare, Kaggle, and BraTs data sets with 98.53% accuracy and 98.47% accuracy in Figshare and Kaggle, respectively, and 90.56% accuracy in the tricky BraTs. While performance is high in terms of generalizability across data sets, drawbacks are that it is computationally expensive and relies on realism of synthetically generated data, which may not always be equal to clinical quality images.

Bouhafra et al. [11] carried out a systematic review of deep learning approaches for brain tumor detection and classification in MRI from 2020 to 2024. The comparison included CNNs, hybrid architectures, transfer learning, and segmentation-assisted models.Their study identified that deep learning techniques tend to surpass conventional hand-engineered features, especially for multi-class classification. They, however, found limitations including reproducibility, usage of dataset variability, and less focus on model interpretability in much of the research.

Filvantorkaman et al. [12] presented a fusion-based approach that combines CNN-based deep learning, explainable AI (XAI), and rule-based reasoning for classifying brain tumors. They combined the CNN learned image features with explainable decision-making reasoning modules. They validated their model on benchmark MRI datasets and achieved > 97% accuracy in glioma, meningioma, and pituitary tumor classification. The innovation lay in combining symbolic reasoning and deep learning but the model was computationally costly and required a lot of preprocessing, which impacted clinic-scale scalability.

Sarker [13] explored transfer learning and explainable AI diagnosis for brain tumor classification from MRI with a Bangladesh dataset. Pretrained models such as VGG16, ResNet, and EfficientNet were fine-tuned and compared. EfficientNet provided the best performance with an overall rate of classification of 96.2%. Grad-CAM was applied to generate visual explanations to decisions that were validated by radiologists to coincide with tumor areas. Although findings demonstrate the promise of transfer learning, dataset size was comparatively small on a proportional basis, restraining reliability within bigger diverse populations.

Yurdakul and Taşdemir [14] proposed a hybrid EfficientNetV2 and MLP-Mixer-Attention brain tumor detector model for MRI. The architecture was developed in such a manner that convolutional efficiency was merged with attention-based global reasoning. The model, which was trained on Kaggle MRI datasets, resulted in 98.7% classification accuracy. Grad-CAM heatmaps were used to provide explainability, although qualitative analysis alone was presented. A major strength was computational efficiency over large Transformer models but the system was not validated using multiple datasets and imaging protocols.

Sharma et al. [15] investigated multi-class classification of MRI brain tumors with information fusion-based hybrid models. They compared CNNs, transfer learning, and ensemble fusion strategies. The optimal performance was achieved using a hybrid fusion model with features of ResNet50 and VGG19 and an overall accuracy of 97.6%. The model was insensitive to inter class variability but involved high feature engineering effort and did not incorporate self supervised learning, thus relying on the availability of labeled data.

Li et al. [16] have explored deep CNN models for brain tumor classification from non-contrast MRI. The work highlighted preprocessing with bias correction and intensity normalization to account for variability in MRI scan. Their proposed CNN model achieved 95.8% accuracy in glioma, meningioma, and pituitary tumor classification. While demonstrating that non-contrast MRI is still able to achieve good classification accuracy, their approach did not include interpretability mechanisms and did not tackle multi-modal imaging.

Gupta et al. [17] used ensemble deep models for brain tumor grading computationally effectively. They used multiple CNNs like DenseNet and Inception as an ensemble for discrimination between low-grade and high-grade glioma. The ensemble average accuracy was 97.1% and F1-score was 96.4%. Ensemble learning performed well in managing minimal grade difference, but the model was computationally intensive and less interpretable, which would decrease clinical adoption.

Khan et al. [18] introduced an EfficientNet-based model (Eff_D_SVM) that combined deep feature extraction with a support vector machine classifier for the classification of brain tumors of diverse types. The model was tested on Kaggle and BraTs datasets with overall accuracy of 96.8%. In novelty, CNN feature extraction was combined with conventional classifiers, enhancing generalizability with small datasets. However, the model needed to be further fine tuned for stable performance across modalities.

Alam et al. [19] introduced a deep ensemble meta-learning platform for the classification of brain tumors. The ensemble made use of several CNN backbones and applied meta-learning techniques to optimization. The platform achieved 99.8% accuracy on 3060 MRI images, surpassing single CNNs. Critically, they implemented Grad-CAM and saliency maps to enhance explainability, and they showed that heatmaps accurately identified tumor areas. Even though performance was robust, the ensemble approach imposed high computational complexity, restricting real-time clinical applications.

Rafiq et al. [20] proposed an ResNet50-augmented classification model using Grad-CAM for interpretable tumor detection. Skull-stripping and data augmentation were incorporated into their preprocessing pipeline. The model was 98.1% accurate, and Grad-CAM highlighted consistency between predictions and tumor positions in MRI scans. The paper highlighted trustworthiness to decision-making but presented low novelty apart from the use of XAI methods within typical CNNs.

Zarenia et al. [21] introduced a deformable attention and saliency-mapping approach for multi-class brain tumor segmentation and classification using MRI. Their MS-DAM module enabled the model to learn how to resolve tumor shape variation with 96.7% accuracy over numerous classes of tumors. Saliency maps were easy to interpret and supplemented segmentation results. Although useful, the model required high-quality annotated segmentation masks, which in practice may prove hard to achieve.

Huang et al. [22] comprehensively reviewed self-supervised learning strategies in medical imaging. Techniques were grouped as contrastive, generative, and masked autoencoders and demonstrated that SSL enormously improved downstream classification precision, especially under low-label scenarios. Although applauding the potential of SSL, the review also identified an absence of standardized metrics, poor evaluation of interpretability, and hurdles in transferring SSL benefits into the clinic.

This paper proposes a powerful deep learning brain tumor classification model from MRI images on the basis of an inferred best SimCLR (SSL) model. The proposed model shows robust performance with 98% test accuracy for four tumor classes, i.e., Glioma, Meningioma, Pituitary, and No Tumor. Significant contributions include a self-supervised contrastive learning model termed CLMC, utilizing unlabeled examples to learn robust feature representations, with additional fine-tuning on labeled data in order to tackle class imbalance as well as MRI scan variability. The model's capability to detect high-level tumor features is confirmed using holistic evaluation metrics such as precision, recall, F1-scores, all above 96%. Grad-CAM visualizations and confusion matrix exploration offer further insight into the diagnostic potential of the model alongside clinical utility. Interpretability was also maximized through usage of explainable AI methods, which showed significant areas in images that help generate predictions. The combination of interpretability with high performance underscores the value of the model for effective application in true real-world clinical practice, particularly where constraints on availability of trained radiologists are present. Future research will target more advancement in precise tumor categorization, integrating multi-modal data such as clinical metadata, and designing light versions for real-time processing to offer best-in-class clinical decision support in various health care environments.

This research utilized publicly available and de-identified MRI datasets, including Figshare, Kaggle, and BraTs. As all datasets are anonymized and contain no personally identifiable information, formal informed consent or ethics approval was not required. The use of these datasets fully complies with relevant institutional, national, and international ethical standards for research involving human data.