Detection of Brain Tumor Employing Residual Network-based Optimized Deep Learning

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Neurosurgery Support

Background:

• Accurate brain tumor diagnosis and treatment planning are crucial for patient survival.
• Tumor variability poses challenges for automated segmentation methods.
• Developing precise automated detection and segmentation techniques for brain tumors is essential.

Purpose of the Study:

• To propose and evaluate novel deep learning models for automatic brain tumor detection and segmentation.
• To improve the accuracy and efficiency of brain tumor analysis in oncological patients.
• To leverage residual networks for enhanced performance in medical image analysis.

Main Methods:

• Utilized a modified ResNet50 model for tumor detection.
• Employed a ResUNet-based convolutional neural network for tumor segmentation.
• Trained and validated models on multi-contrast MRI datasets from 110 patients (TCIA, BraTS2015, BraTS2017).

Main Results:

• Achieved 96.77% accuracy for tumor detection and a 0.893 Dice Similarity Coefficient (DSC) for segmentation on the TCIA dataset.
• Demonstrated improved evaluation parameters, including accuracy and DSC, due to the use of residual networks.
• Validated model performance on BraTS2015 and BraTS2017 datasets, showing consistent and reliable results.

Conclusions:

• Residual networks significantly enhance both detection accuracy and segmentation performance (DSC).
• The proposed models show a 5.9% increase in DSC compared to the standard UNet model.
• Model accuracy improved from 92% to 96.77% on the test set, indicating superior performance.