Sequential LSTM-Based MRI Preprocessing and Attention-Enhanced EfficienNet U-Net for Brain Tumor Segmentation

Abstract

Manual delineation of brain tumors in magnetic resonance imaging (MRI) is time-consuming and subject to inter-reader variability. Automated segmentation can support reproducible tumor delineation, but slice-wise processing does not explicitly represent contextual relationships among adjacent MRI slices. This study investigates a sequential framework that first processes ordered MRI slices with a long short-term memory (LSTM) stage and then performs binary tumor segmentation using an attention-enhanced U-Net with an EfficientNet encoder. Experiments use the public Figshare brain-tumor dataset containing 3,064 T1-weighted contrast-enhanced MRI images from 233 patients with glioma, meningioma, or pituitary tumor. The documented configuration uses RMSprop for the recurrent stage and AdamW for the segmentation network, with batch size 8, a maximum of 500 epochs, dropout 0.5, data augmentation during training, and a Dice-based loss. Performance is assessed primarily with the Dice similarity coefficient (DSC) and Jaccard index (JI). The framework reports a DSC of 92.25% and a JI of 91.41%. The JI value requires verification against the underlying predictions because DSC and JI are mathematically linked for the same binary masks. Literature-based U-Net variants are used only as descriptive reference points because they were not independently retrained under the present protocol. The study therefore provides dataset-specific evidence for the proposed architectural integration while retaining explicit limitations concerning sequence construction, ablation, baseline reproduction, and external validation.