MSB-2.5DUNet++: A Multi-Scale Bottleneck Enhanced 2.5D UNet++ Framework with EfficientNet Encoder for Explainable Brain Tumor Segmentation

Abstract

The segmentation of brain tumors in MRI images still presents challenges because of the heterogeneous appearance of brain tumors, unclear tumor boundaries, and significant class imbalance, which can lead to less accurate and less interpretable automated brain tumor segmentation methods. To overcome these challenges, this study presents an accurate and reproducible framework for brain tumor segmentation in the BRISC 2025 dataset using a modified 2.5D UNet++ model with a multi-scale bottleneck. This study used an EfficientNet-B4 model for brain tumor segmentation in five adjacent axial brain MRI slices at 512 × 512 resolution. This approach enables the model to leverage more contextual information across five slices rather than a single slice. To improve the model's robustness, Kornia-based data augmentation, mixed precision, and exponential moving averages are used. To mitigate the effects of severe class imbalance in brain tumor segmentation, the Dice and Binary Cross-Entropy loss functions are combined. To improve the reliability of brain tumor segmentation, the model uses flip-based augmentation during testing, validation-based threshold sweeping to determine the optimal threshold, and minimum-area filtering to eliminate false positives. In addition, the study used LayerCAM for brain tumor segmentation to improve model interpretability and to develop a CAM-rescue approach for handling empty or very small prediction problems. This study demonstrates the effectiveness of the proposed framework for brain tumor segmentation, achieving Dice scores of 0.87 and 0.862 in validation and testing, respectively, and high AP and AUC values in validation (AP: 0.946, AUC: 0.996) and testing (AP: 0.947, AUC: 0.995). This demonstrates the 2.5D UNet++ model's competitiveness with post-processing for brain tumor segmentation.