Attention-gated U-Net model for semantic segmentation of brain tumors and feature extraction for survival prognosis

TL;DR

This study introduces a tri-planar Attention-Gated Recurrent Residual U-Net (R2U-Net) designed for efficient brain tumor segmentation in MRI and for survival-day (SD) prognosis. Three parallel 2D networks process sagittal, coronal, and axial planes, with their encoder features fused to predict SD through a 3-layer ANN, while segmentation outputs are combined by averaging plane-wise probability maps. The model achieves competitive BraTS2021 validation Dice scores (e.g., WT ≈ 0.900, TC ≈ 0.824, ET ≈ 0.775) and strong specificity, while SD prediction on BraTS2020 yields 45.71% accuracy, MSE ≈ 1.08×10^5, and SpearmanR ≈ 0.338, highlighting both potential and areas for improvement. The approach emphasizes efficiency and interpretability by leveraging multi-plane features and encoder bottlenecks to support clinical decision-making, with future directions including lighter architectures and integration of radiomic and multi-omics data to enhance prognosis accuracy.

Abstract

Gliomas, among the most common primary brain tumors, vary widely in aggressiveness, prognosis, and histology, making treatment challenging due to complex and time-intensive surgical interventions. This study presents an Attention-Gated Recurrent Residual U-Net (R2U-Net) based Triplanar (2.5D) model for improved brain tumor segmentation. The proposed model enhances feature representation and segmentation accuracy by integrating residual, recurrent, and triplanar architectures while maintaining computational efficiency, potentially aiding in better treatment planning. The proposed method achieves a Dice Similarity Score (DSC) of 0.900 for Whole Tumor (WT) segmentation on the BraTS2021 validation set, demonstrating performance comparable to leading models. Additionally, the triplanar network extracts 64 features per planar model for survival days prediction, which are reduced to 28 using an Artificial Neural Network (ANN). This approach achieves an accuracy of 45.71%, a Mean Squared Error (MSE) of 108,318.128, and a Spearman Rank Correlation Coefficient (SRC) of 0.338 on the test dataset.

Figures (6)

• Figure 1: Schematic representation of proposed methodology for brain tumor segmentation and SD prediction. In step 1, a 2D UNet for each axes of the brain MRI is trained to segment the tumor (for each planner view). In step 2, the latent layer of the Encoder part of the 2D UNets (block shaded in green) is used to extract the features. In step 3, these features are analyzed and used to predict the survival days.
• Figure 2: The structural diagram for 2D segmentation network.
• Figure 3: Schematic of the Attention-block used in the 2D UNet.
• Figure 4: CNN network utilized for SD prediction. Here, only the encoder part is used to extract features, where the features are derived from the code layer of the encoder block (as shown in blue.)
• Figure 5: Schematic of Novel SD prediction methodology. The features from each of the encoder blocks of the 2D UNet of the axial, coronal, and sagittal planes are being fused to train an ANN (in orange color blocks) for SD prediction.