A Comparative Study of 2D Image Segmentation Algorithms for Traumatic Brain Lesions Using CT Data from the ProTECTIII Multicenter Clinical Trial

摘要

Automated segmentation of medical imaging is of broad interest to clinicians and machine learning researchersalike. The goal of segmentation is to increase efficiency and simplicity of visualization and quantification of regionsof interest within a medical image. Image segmentation is a di cult task because of multiparametric heterogeneitywithin the images, an obstacle that has proven especially challenging in e orts to automate the segmentation ofbrain lesions from non-contrast head computed tomography (CT). In this research, we have experimented withmultiple available deep learning architectures to segment di erent phenotypes of hemorrhagic lesions found aftermoderate to severe traumatic brain injury (TBI). These include: intraparenchymal hemorrhage (IPH), subduralhematoma (SDH), epidural hematoma (EDH), and traumatic contusions. We were able to achieve an optimalDice Coefficient1 score of 0.94 using Unet++ 2D Architecture with Focal Tversky Loss Function, an increasefrom 0.85 using Unet 2D with Binary Cross-Entropy Loss Function in intraparenchymal hemorrhage (IPH) cases.Furthermore, using the same setting, we were able to achieve the Dice Coe cient score of 0.90 and 0.86 in casesof Extra-Axial bleeds and Traumatic contusions, respectively.