VISION: View-specific integrated segmentation-classification framework for accurate brain tumor detection in MRI scans

Abstract

Brain tumors are an increasing global health concern, and accurate diagnosis is essential for improving patient outcomes. Although existing Magnetic Resonance Imaging (MRI)-based machine learning utilizes computer vision for tumor diagnosis, these methods are limited. They either focus solely on segmentation, which does not facilitate tumor detection, or on classification, which fails to identify tumor boundaries. To overcome these limitations, we propose the “View-specific Integrated Segmentation-Classification” (VISION) framework, designed for more accurate brain tumor diagnosis by integrating both segmentation and classification processes. The VISION framework introduces two novel components: (1) a View Classifier that determines MRI orientation (axial, coronal, or sagittal), and (2) a view-specific integrated network combining a customized segmentation model with a classification header. This architecture simultaneously identifies tumor boundaries (segmentation) and detects tumor presence (classification). We evaluated our approach using publicly available data and compared it against state-of-the-art MRI-based tumor diagnosis techniques. The VISION framework outperformed existing methods, achieving a Dice score of 0.89, an IoU of 0.87, and an F1 score of 0.98 while maintaining competitive computational efficiency. The proposed VISION framework offers a robust solution for brain tumor diagnosis by integrating view classification, segmentation, and detection into a unified system. Its high accuracy and efficiency demonstrate significant potential for clinical applications in improving tumor diagnosis and treatment planning.