Diffeomorphic Registration and Real-Time Intracranial Pressure Estimation Using Hierarchical 3D Vision Transformers

Resumo

A brain tumor pushes healthy tissue aside as it grows. This physical crowding is referred to by doctors as the "mass effect." It is frequently a factor in determining the urgency of surgery and is accompanied by extreme intracranial pressure. Medical AI has mostly focused on segmentation, which is the process of drawing a digital border around a tumor. Finding the bulk is useful, but it ignores the remainder of the brain's mechanical destruction. A fixed form is experienced by the standard segmentation model. It completely ignores the dangerous structural deformation of the surrounding tissue. This study suggests a Deep Volumetric Assessment framework to quantify this invisible pressure. The system determines the exact deviation of the entire brain's physical appearance in addition to describing the illness. It maps a patient's scan to a base of healthy anatomy by mathematically unwarping it using 3D Vision Transformer. Because they operate on discrete patches of pixels, traditional Convolutional Neural Networks (CNNs) are unable to perform such a task. The global geometry is simultaneously used by the Transformer architecture. It tracks the exact mechanism via which a lesion in one hemisphere causes the neuronal connections in the other to lengthen and shorten. The experiment validation on the BraTS dataset shows that this strategy is feasible. The system generates a new Mass Effect Index by generating a high-density tissue motion map. Compared to absolute tumor volume, relative midline shifts are more likely to be predictive of hazardous shifts. By monitoring the physical displacement rather than the physical size as determined by neurosurgeons, this framework provides an automated objective approach of determining the critical brain pressure.