High-resolution segmentation of brain tumors based on MRI images using a hybrid deep learning approach

초록

Today, in decision support systems used by doctors to detect various changes in the brain (tumors, edema, blood clots), magnetic resonance imaging is of great importance, it is a decisive method. However, this method is an invasive method and does not provide high accuracy in many cases. There is a need to develop non-invasive methods to overcome this problem. In this study, brain tumor segmentation was considered. A hybrid deep learning model was proposed as a non-invasive method for brain tumor segmentation. In the developed hybrid model, the ResNext-50 model was selected in convolutional layers to extract many features in the image and reduce gradient degradation from deep learning models, and the U-Net model was selected to increase segmentation accuracy by restoring lost features in the image, and it was named the UnetResNext-50 hybrid model. The dataset in the work was collected from open sources that can be used by scientists and researchers around the world. The dataset used in the study consists of 3929 MRI scans (2556 normal and 1373 tumor images). After initial data processing, including resizing, intensity normalization, and noise removal, the collected data was supplemented with photometric and 12 geometric transformations. As a result, 42,432 samples were obtained. The proposed model was trained with Adam optimizer, binary mutual entropy + dice loss function weights with a batch size of 16. The segmentation performance was evaluated using recall, precision, F1 score, Dice coefficient and Jaccard index. The UnetResNext-50 model achieved Jaccard index of 93.82%, Dice coefficient of 95.7% and overall accuracy of 99.7%. It can be seen from the obtained results that the hybrid model proposed in the study is much more efficient than the traditional deep learning models ResNet-50 and Vanilla U-Net architectures. Quantitative and visual comparisons confirmed that the proposed model provides more accurate tumor boundary detection and better generalization under different MRI contrast conditions. The paper shows that the proposed hybrid model is effective and reliable for real-time medical image segmentation.

키워드