CUDA accelerated Medical Segmentation metrics with MedEval3D

Abstract

Medical segmentation metrics are crucial for development of correct segmentation algorithms in medical imaging domain. In case of three dimensional large arrays representing studies like CT, PET/CT or MRI of critical importance is availability of library implementing high performance metrics. MedEval3D is created in order to fulfill this need thanks to implementation of CUDA acceleration. Most of implemented metrics like Dice coefficient, Jacard coefficient etc. are based on confusion matrix, what enable effective reuse of calculations across multiple metrics improving performance in such use case. Additionally algorithms like interclass correlation and Mahalanobis distance are also introduced. In both cases their implementations are significantly faster then their counterparts from other available libraries. Lastly programming interface to all of the metrics was created in Julia programming language.