Abstract: Manual lesion segmentation for non-contrast computed tomography (NCCT), a common modality for volumetric follow-up assessment of ischemic strokes, is time-consuming and subject to high inter-observer variability. Our approach uses a combination of a 3D convolutional neural network (CNN) combined with post-processing methods. A total of 291 multi-center clinical NCCT datasets were used: 204 for CNN training, 48 for validation and developing post-processing methods, and 39 for testing. The testing datasets were from centers that did not contribute to the training and validation sets, and were segmented by two or three neuroradiologists. We achieved a mean Dice score of 0.42 on the out-of-distribution test set, which was significantly improved to 0.45 with post-processing. The automatically segmented lesion volumes were not significantly different from the lesion volumes determined by manual segmentations from multiple observers. As the model was trained on datasets from multiple centers, it is broadly applicable and publicly available from http://dx.doi.org/10.21227/jps9-0b57 .
Paper Type: both
TL;DR: Development and evaluation of an CNN and post-processing methods for accurate segmentation of stroke lesions in non-contrast tomography images
Track: short paper
Keywords: stroke, computed tomography, segmentation, deep learning, CNN
Presentation Upload Agreement: I agree that my presentation material (videos and slides) will be made public.
Presentation Upload: zip