<p>Deep learning is increasingly being applied across various domains, with health care being one of the most impactful. Despite ethical concerns, its role in rapid and accurate diagnosis, particularly, in computer-aided detection is crucial. Brain tumor classification is a significant challenge within this domain. In this study, we explore transfer learning for brain tumor classification using ResNet18 and ResNet34. Traditional approaches to transfer learning either retrain only final layer or apply optimization-based strategies for selecting trainable layers. We propose an Entropy-Based Fine-Tuning method to selectively retrain layers of a pre-trained network. The approach is grounded in the idea that image entropy reflects information richness; layers with lower entropy are presumed to be underutilized and thus more suitable for fine-tuning. Shannon Entropy is used to quantify this property, and only low-entropy convolutional layers are retrained, rather than the final fully connected layers as in conventional methods. To evaluate the proposed approach, we compare three training strategies: (i) Full Fine-Tuning (all layers trained), (ii) Final Layer Fine-Tuning (only final layer retrained), and (iii) Entropy-Based Fine-Tuning (only low-entropy layers retrained). Two different datasets are analyzed thoroughly. Each model is assessed using accuracy, precision, recall, F1-score, and confusion matrices. Extensive experiments show that the proposed method achieves performance comparable to Full Fine-Tuning while reducing the number of trainable parameters by approximately 10% for ResNet18 and 18% for ResNet34. Moreover, the method demonstrates superior results compared to other studies in this area. Finally, ablation studies confirm that the proposed method is resilient to noise and randomization.</p>