Optimizing glioma vascular analysis: A finetuned masked autoencoder self-attention mechanism approach for accurate...

Accurate segmentation of blood vessels in glioma pathological images is crucial for understanding tumor vasculature and progression, yet remains challenging due to complex vessel morphology and image variations.

This study aimed to develop and evaluate a novel approach using a finetuned masked autoencoder self-attention mechanism (MedSAM) framework for interactive segmentation of glioma blood vessels in pathological images called GliomaVascularSAM.

We utilized a dataset of 2632 image patches derived from multiple glioma datasets. These patches were obtained from tissue samples of 879 patients from The Cancer Genome Atlas and 179 patients from three hospitals. The performance of GliomaVascularSAM was compared with convolutional neural network-based models (including nnU-Net, Pathology-nnU-Net, and nnSAM) and SAM-based segmentation methods. Model performance was evaluated using the dice similarity coefficient, sensitivity, and positive predictive value.

The proposed GliomaVascularSAM outperformed traditional-based models, achieving a dice similarity coefficient of 0.784, sensitivity of 0.767, and positive predictive value of 0.820. Compared with the nnU-Net model (dice similarity coefficient: 0.652), our approach yielded a 13.2% improvement.

GliomaVascularSAM significantly enhanced the accuracy and interactivity of glioma blood vessel segmentation in pathological images. This approach can assist clinicians in the precise analysis of glioma vasculature, thereby contributing to improved diagnosis and management of patients with glioma.