Neural Network-derived perfusion maps: a Model-free approach to computed tomography perfusion in patients with acute ischemic stroke

Abstract

PurposeIn this study we investigate whether a Convolutional Neural Network (CNN) can generate clinically relevant parametric maps from CT perfusion data in a clinical setting of patients with acute ischemic stroke.MethodsTraining of the CNN was done on a subset of 100 perfusion data, while 15 samples were used as validation. All the data used for the training/validation of the network and to generate ground truth (GT) maps, using a state-of-the-art deconvolution-algorithm, were previously pre-processed using a standard pipeline. Validation was carried out through manual segmentation of infarct core and penumbra on both CNN-derived maps and GT maps. Concordance among segmented lesions was assessed using the Dice and the Pearson correlation coefficients across lesion volumes.ResultsMean Dice scores from two different raters and the GT maps were > 0.70 (good-matching). Inter-rater concordance was also high and strong correlation was found between lesion volumes of CNN maps and GT maps (0.99, 0.98).ConclusionOur CNN-based approach generated clinically relevant perfusion maps that are comparable to state-of-the-art perfusion analysis methods based on deconvolution of the data. Moreover, the proposed technique requires less information to estimate the ischemic core and thus might allow the development of novel perfusion protocols with lower radiation dose.

Basis and current state of computed tomography perfusion imaging: a review

“…The second class is to directly estimate HPMs from the low-dose CTP images. For example, Gava et al investigated the potential of the neural networks in HPM estimation from the measured CCTP data, and found that the neural networks can produce clinically relevant HPMs (Gava et al 2021). Robben et al presented a neural network and a data augmentation approach to estimate HPMs directly from the low-dose CTP measurements (Robben and Suetens 2018).…”

“…Clrigues et al proposed a patch based asymmetrical residual encoder-decoder network for infarct core segmentation in the CT/CTP images and accurately predicted infarct core size and location(Clérigues et al 2019). Gava et al constructed a U-Net based network structure to segment infarct core and penumbra with high Dice and Pearson correlation coefficients across lesion volumes(Gava et al 2021). Tao proposed a end-to-end generative model-based segmentation network to segment infarct core and penumbra regions(Tao 2014).…”

UniToBrain Dataset: A Brain Perfusion Dataset

“…Indeed, DL allows the extraction of information and features that are relatively insensitive to noise, misalignment and variance. Gava et al [10] shows whether a properly trained Convolutional Neural Network (CNN), based on a U-Net-like structure, on a pre-processed dataset of CTP images, can generate clinically relevant parametric maps of CBV, CBF and time to peak TTP.…”

“…This particular neural network architecture has been originally developed for image segmentation: however, it proved to be effective also to solve other image generation tasks. For our purpose, we reproduce the suggested modifications [10] at the architectural level to fit the problem. For segmentation tasks, a common state-of-the-art choice is to use max-pooling layers for sub-sampling.…”

“…scan heights or scan angle. A sub-sample of 100 training subjects, and 15 for testing, was used in a submitted publication for the training and the testing of a Convolutional Neural Network[10]…”