Integrating Pyramid Vision Transformer and Topological Data Analysis for Brain Tumor

Dhananjay Laxmikant Joshi ^1* Bhupesh Kumar Singh ¹ Kapil Kumar ² Nitin S Choubey ³

• ¹ Amity University Jaipur, Jaipur, India
• ² Guru Ghasidas Vishwavidyalaya, Bilaspur, Chhattisgarh, India
• ³ SVKM's Narsee Monjee Institute of Management Studies, Mumbai, Maharashtra, India

Accurate Brain Tumor (BT) classification is crucial due to the complex and varied nature of tumors. This paper presents a novel approach for BT classification that combines a Pyramid Vision Transformer (PVT) with an adaptive deformable attention mechanism and Topological Data Analysis (TDA) to address the complexities of BT detection. While PVT and deformable attention mechanisms have been explored in prior works, this study introduces several key innovations to enhance their performance for medical image analysis. Specifically, we propose an adaptive deformable attention mechanism that dynamically adjusts receptive fields based on tumor complexity, allowing for better focus on critical regions in MRI images. Additionally, we introduce an adaptive sampling rate with hierarchical dynamic position embeddings, enabling context-aware feature extraction across multiple scales. The framework further improves feature diversity through an offset group mechanism that partitions feature channels into specialized groups. To integrate local and global contexts, we employ a hierarchical deformable attention strategy, refining the feature representation. Topological insights from TDA, applied to preprocessed images, facilitate the extraction of meaningful patterns for classification. The BT classification is performed using a Random Forest Classifier. Extensive testing on the Figshare database demonstrates the effectiveness of the proposed methodology, yielding an accuracy of 99.2%, recall of 99.35%, precision of 98.9%, F1-score of 99.12%, Mathews Correlation Coefficient (MCC) of 0.98, LogLoss of 0.05 and an execution time of just 6 seconds. These results highlight the method's ability to combine detailed feature extraction with topological insights, significantly advancing the accuracy and efficiency of BT classification and offering a promising tool for improving diagnostic outcomes.