AI Deep Learning Methods for Segmenting 3D-Brain Tumors Images using U-Net

Deep learning approaches have transformed medical image segmentation, particularly for the automatic separation of brain tumours in pre-operative MRI scans. This study suggests a better U-Net design to fix the problems with current ones, like not splitting edge features enough, not getting location information well, and using feature details too many times without changing them, which are all problems with older models. The suggested model combines a serial encoding-decoding method with a hybrid dilated convolution module. It also includes inter-module concatenations to improve the accuracy of segmentation. To improve classification and segmentation performance, a new loss function that prioritizes difficult-to-segment data is also developed. The BraTS dataset, which includes MRI images of different types (T1, T1c, T2, and FLAIR) found in three different parts of brain tumors—enhanced tumor (EST), tumor core (TC), and entire tumor (WT)—was used to test how well the model worked. The segmentation results are judged using metrics such as dice, an IOU, Hausdorff95, and PA among ASD. These show that the suggested model consistently works better than current methods. Visual comparisons provide additional confirmation of its capacity to capture more detailed tumour features and generate results that are more accurate. The results show that the model is very good at semantic segmentation, which could have effects on clinical workflows like therapy planning. They also show that the model can be used to help with brain tumor research for both diagnosis and treatment. The proposed model achieved superior segmentation performance with an mIoU of 89.45%, MPA of 92.35%, precision of 95.98%, and Dice of 94.01%. These results demonstrate its strong capability in accurately segmenting brain tumours, outperforming existing methods.