FDS-ConvLSTM U-Net: A Sequence U-Net for Brain MRI Segmentation by Four-Directions Scan Conv-LSTM

Since the beginning of this year, the trend of combining image segmentation with sequence models like Transformers has gained significant popularity, leading to remarkable advancements in deep learning in this field. However, effectively leveraging these sequence models remains a challenge. To address this, we propose a novel medical image segmentation framework, FDS-ConvLSTM U-Net, which skillfully combines the strengths of Convolutional Long Short-Term Memory Networks (Conv-LSTM) with the U-Net architecture. To enhance segmentation performance further, we introduced an innovative Four-Direction-Scan (FDS) strategy, which leverages a self-attention mechanism to precise local features and distant contextual relationships from multiple directions. On the BraTS2023 dataset, Dice Similarity Coefficients (DSC) reached 0.9276, 0.9275, and 0.8710 for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET), respectively, using the FDS-ConvLSTM model.