OPTIMIZED LSTM-DRIVEN MRI PREPROCESSING WITH FINE-TUNED U-NET FOR ACCURATE BRAIN TUMOR SEGMENTATION

The evaluation of brain tumors within MRI scans stands as a fundamental medical imaging operation which benefits diagnostic procedures as well as treatment arrangement processes. The research investigates an optimized arrangement of Long Short-Term Memory (LSTM) networks that process MRI scans before using a modified U-Net network for brain tumor segmentation. The sequential LSTM model processes MRI sequence images so it improves spatial coherence and reduces noise to enhance image quality by adapting intensity normalization. The initial processing operation enhances the quality of the input data thereby improving its readiness for segmentation. A U-Net architecture with fine-tuned capabilities uses EfficientNet as its backbone in the encoder section while adding attention mechanics to the decoder framework. The model optimization strategy includes the combination of Dice and Focal losses for effective class imbalance management. The model performance benefits from additional improvements which come from both data augmentation methods and meticulous hyperparameter adjustments. Research experiments performed on standard MRI datasets show that the proposed method outperforms past methodologies at achieving higher segmentation outcomes. The proposed method attaining 10.5%, 7.11%, 1.26 improvement with reference to Dice Similarity Coefficient (DSC), 8.56, 6.91, 2.36 improvement with reference to Jaccard Index (JI) and 10.23, 5.69, 3.94 improvement in segmentation accuracy over the U-Net, Attention U- Net, Res Attention U-Net models. The advanced U-Net architecture together with sequential LSTM preprocessing allows the system to produce both effective feature detection and accurate tumor outline identification. The approach demonstrates prospective capability to support radiologists in their automated diagnosis of tumors and therapeutic planning activities. The authors will analyze real-time implementation and clinical application optimization during subsequent stages of their research.