Emirhan Kurtulus

Society for Science & the Public (ISEF)

DEEP LEARNING-BASED STEREOTACTIC CRANIAL SURGERY PLANNING

 

Cancer and cranial diseases are the second leading cause of death worldwide. Early diagnosis and well-guided medical care as well as accurate analysis are essential to improve patient life expectancy and quality. We propose the first endto- end deep learning pipeline for calculating cranial surgery risk factors to plan with minimal neurological damage. The proposed system is able to plan cranial surgeries using MRI scans of the patient and consists of a new scalable state-of-the-art (+3.2%) UNet variant based on depthwise separable convolutions for brain tumor segmentation; a new state-of-the-art (+2.1%) transformer-based architecture for tractography; a new state-of-the-art (+3.6%) anatomical brain segmentation architecture orders of magnitude faster and smaller; a state-of-the-art tumor classification model (+4.3%) trained with a new data augmentation pipeline, and a novel ML algorithm that learns the importance of anatomical regions for planning the surgery with the outputs of these models. The performance of the system is validated by a retrospective clinical study. It is shown to be able to detect the tracts to be damaged during surgery with 96.12% accuracy. We have also developed compact desktop and augmented reality applications that can be used in well-equipped medical centers as well as rural hospitals.