Comparing deep learning stroke segmentation in NCCT, CTA, and CTP: Accuracy, domain transfer, and temporal sampling effect

Abstract

Background

Stroke imaging typically involves multiple CT image types—non-contrast CT (NCCT), CT angiography (CTA), and CT perfusion (CTP). CTP and multiphase CTA (mCTA) are more advanced acquisitions with multiple timesteps and provide insights on the hemodynamics within the brain. Deep Learning models can help facilitate the diagnostic workflow by automatically identifying the extent of core and penumbra, which influences subsequent treatment decisions. For the use in clinical practice, generalizability of these models to new clinical sites is crucial.

Purpose

We evaluate and compare the usefulness of NCCT, CTA, mCTA, and CTP images for DL-based stroke lesion segmentation, with the aim of guiding modality selection in settings with and without access to advanced imaging, and with an additional focus on model transferability between clinical sites and the impact of time point selection from the CTP scan.

Methods

The experiments involve model training with a dataset of 91 stroke patients from one clinical site. NCCT, CTA, mCTA, and CTP are used separately to train nnU-Net models for segmentation of stroke core and hypoperfused volume using uncertainty-aware labels. To assess site transferability, a model (pre-)trained on 166 cases from a second clinical site is employed to perform as-is inference with data from the first site, then contrast it with a variant of the model fine-tuned using a subset of the data from the first site. Multiple temporal sampling strategies were investigated for the 4D CTP data, choosing different subsets of the time series as the model input.

Results

For automatic segmentation of stroke core, advanced imaging techniques yield improved accuracy with the modified Dice coefficient increasing from $0.36\pm 0.28$ (NCCT) to $0.55\pm 0.27$ (CTA), $0.71\pm 0.22$ (mCTA), and $0.78\pm 0.09$ (CTP) for infarcts of size 10–70 mL. A similar trend is observed for smaller infarcts of 1–10 mL. In terms of generalizability, the additional fine-tuning stage consistently enhances the segmentation results, regardless of the image type used. To leverage the initially large series of perfusion images, different temporal sampling strategies are applied to predict stroke core. The experiments show no clear trend as the results vary across different timing scenarios and infarct sizes.

Conclusions

The study provides an overview of the quality of automated stroke lesion segmentation with nnU-Net across all relevant CT acquisition types. Hereby, multitimepoint imaging exhibits significantly improved segmentation performance compared to NCCT and CTA.