ABSTRACT Brain tumour classification using convolutional neural networks (CNNs) has emerged as a vital application in medical image analysis. Optimizing the CNN architecture and removing unnecessary neurons play pivotal roles in enhancing the model's performance for accurate tumour identification. This article presents a comprehensive exploration of techniques to improve CNN efficiency for brain tumour classification in low‐resolution magnetic resonance imaging (MRI). The study delves into pruning the CNN architecture and leveraging transfer learning to bolster the model's capabilities. During the neuron pruning, hyperparameter tuning and regularization are performed to reduce model complexity and prevent overfitting. By incorporating techniques such as increased dropout rates and early stopping, significant improvements are achieved in model performance. Evaluation metrics and performance analysis show the optimized CNN's superiority over the baseline model. The dataset employed in this study consists of 658 MRI images, organized into two classes: 483 tumour‐containing cases and 175 non‐tumour cases. The proposed method demonstrates a mini‐batch accuracy increase from 66.41% to 97.66% within 50 iterations and achieves a low mini‐batch loss of 0.0099 at 100 iterations. Due to its lightweight architecture and reduced computational complexity, the proposed model is well‐suited for practical clinical deployment, particularly in resource‐constrained diagnostic environments.
Afsarinejad et al. (Thu,) studied this question.