Attention-enhanced SAM with PBFO tuning: advancing glioma MRI segmentation

TL;DR

This paper introduces PoSAM-ULTRA, a new framework for brain tumor MRI segmentation that improves accuracy and robustness using attention mechanisms and a novel optimization algorithm.

Contribution

The novel contribution is PoSAM-ULTRA, which combines an improved Segment Anything Model with PBFO optimization and attention modules for enhanced brain tumor segmentation.

Findings

PoSAM-ULTRA achieved a Dice score of 91.4%, outperforming UNet, UNet++, and nnUNet.

The model demonstrated high accuracy (99.8%) and robustness in handling complex brain tumor MRI data.

Abstract

The segmentation of brain tumor MRI images is one of the most challenging tasks because of the variability and complexity associated with tumor tissues. This study introduces PoSAM-ULTRA, an improved segmentation framework designed to enhance the accuracy and robustness of brain tumor segmentation.

PoSAM-ULTRA employs the Polar-Bear Foraging Optimisation (PBFO) algorithm for hyperparameter tuning and utilizes an improved Segment Anything Model as its backbone. The framework is based on a ResNet-34 encoder modified to accept a four-channel input (RGB + prior information). Multi-scale feature extraction is performed via DownBlocks, while discriminative feature learning is enhanced using the Convolutional Block Attention Module (CBAM). Attention Gates are incorporated to ensure effective skip connections, and a multistage decoder is used for robust upsampling and feature integration. The model was evaluated on a dataset from the Integrative Genomic Analysis of Diffuse Lower Grade Gliomas (LGG) and compared with UNet, UNet++, and nnUNet.

The proposed PoSAM-ULTRA model outperformed the baseline models, achieving superior performance with a Dice score of 91.4%, IoU of 88.9%, Accuracy of 99.8%, Precision of 95.2%, and Recall of 93.3%.

The obtained results demonstrate the robustness and reliability of PoSAM-ULTRA in handling the complexity of brain tumor MRI segmentation, highlighting its effectiveness for challenging medical image segmentation tasks.

Figures