Related Experiment Video
Updated: Jan 8, 2026
Author Spotlight: AI-Driven Trypanosome Species Detection from Microscopic Images
Published on: October 27, 2023
Hybrid Deep Learning for Brain Tumor Detection: Combining DenseNet and Custom CNN for Enhanced Accuracy
Alex David Swaminathan1, Almas Begum2, Karthikeyan Ramamoorthy3
1Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai, Tamilnadu, India.
Background:
Deep learning in brain tumor detection has become an important breakthrough in medical imaging to facilitate a fast and accurate diagnosis. Conventional models such as VGG, SVM, and common CNNs are competitive, yet fail to provide the sensitivity and specificity needed in real-time diagnosis purposes.
Objectives:
The objective of the proposed study is to generate a hybrid deep learning architecture of the DenseNet and a self-developed Convolutional Neural Network (CNN) in order to increase the classification accuracy, sensitivity, and specificity of a brain tumor detection medical image.
Design:
A Hybrid architecture consisting of DenseNet feature reuse and connectivity with a domain-specific custom CNN is suggested to retrieve high-level semantic and fine-grained features. The design focuses on performance evaluation with respect to the state-of-the-art models and ensemble frameworks.
Methods:
The brain tumor data set was preprocessed, and the augmentation methods, such as translation and rotation, were applied. A training and testing subsets of the dataset were formed. The hybrid approach, formed by DenseNet layers and a self-created CNN model, was implemented and tested. The performance of the model was contrasted with that of other benchmark classifiers such as SVM and VGG, DenseNet (single), CNN ensembles, and Hybrid Ensembles.
Results:
The hybrid DenseNet-Custom CNN model was more accurate and had better classification results than the standard models. In particular, it surpassed the accuracy of SVM (96%), VGG (94%), DenseNet (92%), and Hybrid Ensemble models (~95.2%), while the FPS remained equivalent to SVM and substantially lower than VGG. It proved better sensitivity and specificity with better feature representation and interpretation, as it made more accurate tumor classification.
Conclusion:
Incorporating DenseNet with a custom CNN model can increase the capabilities of brain tumor detection in medical imaging. This hybrid performance can use both general-purpose deep learning and domain-specific feature engineering, and it provides a practical suggestion of the approach that involves a satisfactory solution in the diagnostic sense. It verifies that the combination of more than 2 methods is productive in enhancing the results of medical image classification activities.