Brain tumor segmentation in three-dimensional mri data using a hybrid deep learning approach based on attention mechanisms

The purpose of this study is about a new detailed model that is designed to perform brain tumor segmentation on three-dimensional magnetic resonance imaging (3D MRI) data. The model is a combination of both Gated Attention and CBAM structures to more effectively capture both types of data. The model is built on a classical encoder-decoder structure. However, a Gated Attention layer is placed in the transition region between the encoder and decoder, and CBAM blocks are unified into skipping the connections of the decoder part. The model was trained and evaluated with careful separation on the BraTS 2020 dataset. The average performance achieved on the test data is as follows: accuracy 0.9934 ± 0.0034, Dice coefficient 0.8768 ± 0.0738, IoU score 0.7876 ± 0.1079, and recall 0.8409 ± 0.1192. Therefore the difficulty of the model was measured at 83.32 GFLOPs, and it has showed amazing efficiency with only 0.98M parameters. The results show that the dual attention mechanism structure offers a competitive and balanced performance compared to many methods in the literature, both in terms of overall accuracy and success in identifying tumor boundaries.