nnU-Net Segmentation for the Optimization of Treatment and Outcome Prediction After Subarachnoid Hemorrhages

Abstract

Accurate quantification of initial subarachnoid hematoma volume (ISHV) is crucial for evaluating the severity of aneurysmal subarachnoid hemorrhage (aSAH). Traditional manual segmentation methods on non-contrast CT (NCCT) are complex, time consuming, and prone to errors specially when performed by less experienced clinicians. This project explores the use of convolutional neural networks (CNN) for the automatic segmentation of ISHV. The nnU-Net algorithm was employed, conducting a series of experiments to enhance its efficiency and accuracy. We investigated the influence of pre-processing input scans with a brain extraction algorithm prior to training, and evaluated the nnU-Net robustness by training and testing with different data splits. Additionally, statistical analysis were conducted to study the differences between experiments and their corresponding influence on the performance of the trainings. Results showed an improvement in performance using the nnU-Net, obtaining a median Dice score of 0.76 and a median 95\% Hausdorff distance of 9.1 mm. We also demonstrated a slight upgrade when including the custom pre-processing pipeline, as well as proving the importance of data splitting. However, the statistical analysis did not reveal significant differences in most of the experiment comparisons. Overall, we provide a robust foundation for future advancements, revealing promising preliminary findings and highlighting crucial aspects of aSAH diagnosis that could be important for the ongoing research of the topic.