Precision meets generalization: Enhancing brain tumor classification via pretrained DenseNet with global average...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Oncology

Background:

• Brain tumors present a significant global health challenge with high mortality rates.
• Manual detection of brain tumors from MRI scans is subjective and difficult.
• Automated solutions are crucial for improving diagnostic accuracy and efficiency.

Purpose of the Study:

• To investigate the efficacy of deep learning models for automated brain tumor classification.
• To compare the performance of DenseNet against other architectures like ResNet, EfficientNet, and MobileNet.
• To enhance the accuracy and generalizability of a DenseNet model for clinical applications.

Main Methods:

• Utilized the Figshare brain tumor dataset of 3,064 T1-weighted contrast-enhanced MRI images.
• Evaluated four pre-trained deep learning models (ResNet, EfficientNet, MobileNet, DenseNet) using transfer learning from ImageNet.
• Implemented fine-tuning with regularization techniques (data augmentation, dropout, batch normalization, global average pooling) and hyperparameter optimization on DenseNet.

Main Results:

• DenseNet achieved the highest initial test accuracy at 96%, outperforming ResNet (91%), EfficientNet (91%), and MobileNet (93%).
• The fine-tuned DenseNet model demonstrated improved performance, reaching an accuracy of 97.1%.
• The study highlights the effectiveness of transfer learning and fine-tuning strategies.

Conclusions:

• DenseNet, enhanced through transfer learning and fine-tuning, is a highly effective tool for brain tumor classification.
• The proposed method shows significant potential for improving diagnostic accuracy and reliability in clinical settings.
• Automated classification using deep learning can aid healthcare professionals in brain tumor diagnosis.