An Image Analysis Method for Perihematomal Edema in Cranial CT Images Based on Deep Learning

Intracerebral hemorrhage (ICH) has become the top cause of mortality from worldwide diseases. Meanwhile, spontaneous intracerebral hemorrhage (SICH), which is usually caused by a burst or abnormal vessel inside brain tissue, tends to lead to bad prognostic outcomes. Perihematomal edema (PHE), resulting from SICH, has been regarded as an important matter to be detected and segmented as soon as possible for better ICH patients’ prognosis but was always postponed due to lack of public PHE’s dataset up till now and low accuracy & efficiency of traditional machine learning & deep learning approaches caused by overlapped density & morphology among PHE images. Therefore, we establish our own PHE-SICH-CT-IDS including 120-brain CTs image & 7,022-images which can provide us a perfect base for subsequent PHE segmentation, detection, radiomics study. And we design new SMS-UNet model with state space & dual cross attention mechanism introduced for further increasing the segmentation’s accuracy & efficiency. In conclusion, we successfully built a novel & large-scale dataset based on patient’s SICH-CT images & corresponding PHE masks; besides, the experiment results showed that the proposed dataset is suitable for testing different kinds of methods in segmentation, detection and extraction of diagnostic radiomics features. Finally, we proved through multiple evaluation indicators that SMS-UNet model have obtained extremely good results, especially dice coefficient. Considering that PHE-SICH-CT-IDS becomes the first publicly shared data on this aspect in whole academic circle, therefore, the establishment of such database may promote diagnosis and treatment of ICH with high contribution.