Advanced Fusion of 3D U-Net-LSTM Models for Accurate Brain Tumor Segmentation

Abstract

Accurate detection and segmentation of brain tumors are essential in tomography for effective diagnosis and treatment planning. This study presents advancements in 3D segmentation techniques using data from the Kaggle BRATS 2020 dataset. To enhance the reliability of brain tumor diagnosis, innovative approaches such as Frost filter-based preprocessing, UNet segmentation architecture, and Long Short-Term Memory (LSTM) segmentation are employed. The methodology starts with data preprocessing using the Frost filter, which effectively reduces noise and enhances image clarity, thus improving segmentation accuracy. Subsequently, the UNet architecture is utilized to precisely segment brain tumor regions. UNet's ability to capture contextual information and its efficient use of skip connections contribute to accurately delineating tumor boundaries in three-dimensional space. Additionally, the temporal aspect of brain tumor progression is addressed by employing an LSTM network, which increases segmentation accuracy. The LSTM algorithm integrates temporal patterns in sequential imaging data, enabling reliable segmentation of tumor presence and characteristics over time. By analyzing the ordered sequence of continuous MRI scans, the LSTM framework achieves more precise and adaptable tumor recognition. Evaluation results based on the Kaggle BRATS 2020 dataset demonstrate significant improvements in segmentation and segmentation performance compared to previous methods. The proposed approach enhances the accuracy of tumor boundary delineation and the ability to classify tumor types and track temporal changes in tumor growth. The "U-Net-LSTM" method achieves an accuracy of 98.9% in segmentation tasks, showcasing its superior performance compared to other techniques. This method is implemented using Python, underscoring its efficacy in achieving high accuracy in segmentation tasks.