CerevianNet: Parameter Efficient Multiclass Brain Tumor Classification Using Custom Lightweight CNN

Md Khurshid Jahan¹Abdullah Al Shafi¹Rusho Maher Ali^2*Md. Shahriar Hussain¹

• ¹North South University, Dhaka, Bangladesh
• ²University of Colorado Boulder, Boulder, United States

Brain tumors are a life-threatening condition, and their early detection is crucial for effective treatment and improved survival rates. Traditional manual evaluation techniques, such as expert radiologist assessments and visual inspections, are widely used for diagnosing brain tumors. While these methods can be highly reliable, they are often time-consuming, prone to human error, and challenging to scale for large datasets. Consequently, there is a growing demand for computer-aided diagnostic systems to overcome these limitations and deliver fast, accurate, and scalable solutions. Despite these promising advancements, the study highlights potential limitations, such as susceptibility to overfitting due to limited labeled data and the need for extensive hyperparameter tuning to generalize across diverse datasets. This paper proposes a scalable multi-class brain tumor classification framework optimized for small-form-factor devices. We introduced a novel lightweight custom convolutional neural network (CNN) that maintains high classification accuracy while significantly reducing computational complexity. We evaluated the capacity of the model by training and testing on five different datasets and the model performed well on all five. We saw great performance with the model on larger datasets, but it struggled when it came to smaller and imbalanced datasets. We achieved significant scores on the datasets, and we had the highest testing accuracy on Dataset-5 (99.67% training accuracy, 98.17% validation accuracy and 98.30% testing accuracy). What is important to note is that we had the lowest testing accuracy on Dataset-3 (99.99% training accuracy, 74.11% validation accuracy and 75.63% testing accuracy).. The proposed framework leverages state-of-the-art pre-trained deep learning models, including EfficientNetb3, ResNet-101, ResNet-50, Xception, AlexNet, DenseNet121, Swin Transformer, and our custom lightweight CNN model. Experimental evaluations demonstrate that EfficientNetb3 achieves the highest accuracy of 99.11%, while the custom lightweight CNN attains 98% accuracy with 4.1x fewer parameters and reduced training time. These results highlight the effectiveness of computer-aided approaches in achieving near-expert performance, making them suitable for integration into clinical workflows. This research paves the way for deploying efficient and scalable deep learning models in real-world medical applications to expand accessibility to accurate brain tumor diagnosis.