Random expert sampling for deep learning segmentation of acute ischemic stroke on non-contrast CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate delineation of acutely infarcted brain tissue on non-contrast CT is crucial but challenging due to limited inter-reader agreement among human experts.
• Supervised deep learning offers a potential solution, but optimal training strategies for segmenting ischemic tissue remain under investigation.

Purpose of the Study:

• To compare the performance of two deep learning training methods for segmenting ischemic brain tissue on non-contrast CT: majority vote from expert segmentations versus random sampling of individual expert segmentations.
• To evaluate the agreement of the deep learning models with expert consensus and with 24-hour follow-up diffusion-weighted imaging (DWI) infarct core volumes.
• To assess the correlation of model-derived infarct volumes with clinical outcomes.

Main Methods:

• A U-Net deep learning model was trained using data from 260 non-contrast CT studies of acute ischemic stroke patients (DEFUSE 3 trial) and validated with 33 external cases.
• Two training schemes were employed: (1) majority vote from three expert neuroradiologists' segmentations, and (2) random sampling from individual expert segmentations.
• Segmentation performance was assessed using volume, overlap (Dice score), and distance metrics, compared against inter-expert agreement and each other using Wilcoxon signed rank tests.

Main Results:

• The random expert sampling model achieved significantly higher Dice scores (0.51±0.04) compared to both inter-expert agreement (0.36±0.05) and the majority vote model (0.45±0.05).
• The random model's predicted infarct volume showed a significant correlation with 90-day modified Rankin Scale (mRS) scores, unlike the median expert or majority models.
• No significant differences were observed in volume correlations between the models and the 24-hour follow-up DWI infarct core volume.

Conclusions:

• A deep learning model trained on random expert segmentations demonstrates superior performance in delineating ischemic injury on non-contrast CT compared to majority vote and inter-expert agreement.
• The volumetric measurements from the random model are consistent with final infarct core volumes determined by 24-hour follow-up DWI and correlate with clinical outcomes.