Brain tumor classification from magnetic resonance imaging (MRI) is a common direction in the application of deep learning to medical image analysis. However, this task often faces a significant challenge: class imbalance. Some tumor types appear frequently in datasets, while others are rare, which can lead to models that are biased-overfitting on the more prevalent classes and struggling with underrepresented ones. In this study, we introduce a method that modifies the standard CrossEntropy loss function by incorporating class weights based on the concept of effective sample size. Instead of simply using raw sample counts to adjust for imbalance, our approach attempts to better reflect the actual contribution of each class during training. This helps mitigate the model’s tendency to favor dominant classes. We tested this approach on two modern deep neural network architectures, aiming to evaluate its ability to enhance classification performance under label imbalance—without altering model structures or relying on resampling strategies. The goal is to offer a lightweight, adaptable solution that can be easily integrated into existing medical image analysis pipelines.