ENHANCED BRAIN TUMOR SEGMENTATION WITH SKIP CONNECTED RESNET-18 ENCODER-DECODER FRAME WORK WITH DICE LOSS

The research develops an advanced brain tumor segmentation system that combines a skip connection between ResNet-18 and Dice Loss in an encoder-decoder network. Brain tumor segmentation helps medical imaging specialists find and plan better treatment methods for patients. The proposed method uses ResNet-18 as a base which combines deep architecture with residual connections to prevent vanishing gradient problems. In the technique the model connects the encoder and decoder to keep spatial details that disappear in typical architectures during down sampling steps. The method can better identify hard-to-sector tumor areas because it handles small and nonuniform tissue elements effectively. The model uses Dice Loss as its optimization method to measure how well the predicted tumor outlines match the actual tumor areas. The proposed network proves better tumor segmentation results on its regular brain data collection through testing against older technique approaches. The new model outperforms U-Net by 0.87% in accuracy plus 46.99% in precision along with better F1 scores and Dice coefficients but demonstrates an 0.84% drop in recall. The new method proves better than U-Net by revealing improved performance and segmentation outcomes. The network design efficiently segments challenging areas using its encoder-decoder structure based on ResNet-18 and Dice Loss calculations.