CerevianNet: parameter efficient multi-class brain tumor classification using custom lightweight CNN

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Traditional manual brain tumor diagnosis is time-consuming and prone to errors.
• Computer-Aided Diagnostic (CAD) systems offer faster, scalable solutions.
• Deep learning models face challenges like overfitting with limited data.

Purpose of the Study:

• To propose a scalable multi-class brain tumor classification framework for small-form-factor devices.
• To develop a lightweight custom convolutional neural network (CNN) for efficient brain tumor diagnosis.
• To evaluate the performance of the custom CNN against state-of-the-art deep learning models.

Main Methods:

• Developed a novel, lightweight custom convolutional neural network (CNN).
• Evaluated the custom CNN and pretrained models (EfficientNetb3, ResNet, etc.) on five diverse brain tumor datasets.
• Optimized the framework for small-form-factor devices and assessed performance on varying dataset sizes and balances.

Main Results:

• The custom lightweight CNN achieved 98% accuracy with significantly fewer parameters and reduced training time compared to other models.
• EfficientNetb3 demonstrated the highest accuracy at 99.11%.
• The model performed well on larger datasets but struggled with smaller, imbalanced ones, highlighting data dependency.

Conclusions:

• The proposed framework effectively utilizes deep learning for accurate brain tumor classification, approaching expert performance.
• The lightweight custom CNN offers an efficient and scalable solution suitable for clinical integration.
• This research facilitates the deployment of AI in medical applications for improved brain tumor diagnosis accessibility.