Predicting hematoma expansion in intracerebral hemorrhage from brain CT scans via K-nearest neighbors matting and deep residual network

Hematoma Expansion (HE) in spontaneous IntraCerebral Hemorrhage (ICH) is one of the highest mortality rates in neurosurgery. It is also influenced by various factors such as hematoma volume, location, shape, and so on. Although it’s crucial to present a method to predict HE, there is no work on the prediction model using brain imaging for this task to our knowledge. Inspired by the above, this work proposes a method including preprocessing and classification to predict HE in ICH from brain computed tomography (CT) scans. A k-nearest neighbors matting method is adopted in the preprocessing of brain CT scans to remove the outer part of the skull and retain brain tissue features. A deep residual network is then presented to give classification results, which helps to learn from features of hematoma and other parts in brain tissue. Experimental results on 223 patients including 137 HE patients show that the proposed framework can achieve 0.890 ± 0.020 accuracy, 0.880 ± 0.033 specificity, 0.925 ± 0.041 sensitivity, and 0.867 ± 0.025 F1-Score. Compared with other state-of-the-art works that need a large number of clinic data, the proposed method can obtain better prediction performance using brain CT scans only.