An intelligent LinkNet-34 model with EfficientNetB7 encoder for semantic segmentation of brain tumor

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Neurology

Background:

• Brain tumors are aggressive, necessitating accurate segmentation for early diagnosis.
• Manual segmentation by radiologists is time-consuming, labor-intensive, and prone to errors.
• Deep learning models demonstrate potential to surpass human expert performance in brain tumor diagnosis.

Purpose of the Study:

• To propose an automated semantic segmentation model for brain tumors in MRI images.
• To enhance the accuracy and efficiency of brain tumor segmentation using deep learning.
• To evaluate the proposed model's performance against existing methods and various configurations.

Main Methods:

• An encoder-decoder architecture utilizing a deep convolutional neural network was developed.
• The proposed model, LinkNet-34, incorporates an EfficientNetB7 encoder, focusing on image downsampling.
• Performance was benchmarked against FPN, U-Net, and PSPNet, with encoder variations (ResNet34, MobileNet_V2, ResNet50) and optimizers (RMSProp, Adamax, Adam) tested.

Main Results:

• The LinkNet-34 model with the EfficientNetB7 encoder demonstrated superior performance.
• Optimization with the Adamax optimizer yielded the best results.
• The model achieved a Jaccard index of 0.89 and a Dice coefficient of 0.915.

Conclusions:

• The proposed LinkNet-34 model with EfficientNetB7 encoder offers a highly accurate solution for brain tumor segmentation.
• The Adamax optimizer significantly enhances the model's segmentation capabilities.
• This deep learning approach promises improved efficiency and accuracy in medical image analysis for brain tumor diagnosis.