Usefulness of Monoenergetic Non-contrast CT and X-Map Images for Deep Learning-Based Stroke Lesion Segmentation

Abstract

This study investigates the usefulness of monoenergetic non-contrast computed tomography (NCCT) and x-map images in comparison to conventional NCCT for neural network-based stroke lesion segmentation. Utilizing the nnU-Net segmentation framework, models are trained on conventional, Mono+50, Mono+70, and Mono+120 NCCT as well as x-map images. Performance is evaluated on all image types, including those not seen during training, with the network predicting stroke core and hypoperfused volumes. Evaluation metrics include Dice scores and volumetric measurements. Results indicate that training nnU-Net with Mono+120 images on average yields the best performance across all tested image types. While x-map images, previously shown to facilitate lesion detection in human reader studies, do not outperform Mono+120 in same-domain training and testing, they demonstrate robust performance in inference for different energy level and conventional NCCT-trained models. For medium sized infarcts of 10–70 ml, the best Dice score is achieved when training with Mono+120 and testing on x-map images. These findings highlight the potential benefits of advanced spectral CT image derivatives for ischemic stroke segmentation.