Hemorrhagic transformation (HT) is regarded as a safety endpoint of arterial ischemic stroke acute treatment and secondary prevention trials (Hutchinson and Beslow, 2019). Accurate HT prediction dramatically reduces the death rate from misdiagnosis. At present, HT predictions almost all rely on contrast images with perfusion agents, which is time-consuming and labor-intensive, causing secondary brain damage and high cost. Almost all machine learning algorithms cannot use non-contrast CT for HT prediction because of huge challenges. In this study, a Dual-branch Separation and Enhancement Network (DBSE-Net) is proposed for weak feature extraction and safe HT prediction without perfusion agents. DBSE-Net innovatively uses a dual-branch separation and fusion mechanism to achieve weak feature adaptive extraction. In the DBSE-Net’s encoder submodules, Brain Compression Assessment Branch (BCAB) and Infarct Assessment Branch (IAB) are proposed to apply lightweight encoding structures with different receptive fields, which are adapted to the lesion area’s characteristics. With the help of DBSE-Net’s keyframe selection algorithm and area guidance knowledge, DBSE-Net removes redundant information and clearly describes the severity of lesions. In summary, DBSE-Net integrates global and local features to obtain multi-scale and multi-category brain status information, enhancing the weak features of non-contrast CT and realizes accurate HT prediction. Experimental Result: Among all 144 intracranial stroke patients diagnosed by doctors as having no HT risk, DBSE-Net identified 73 high-risk HT patients (88 HT cases in total). The result illustrates that DBSE-Net helps doctors secondary diagnose the HT risk of intracranial stroke patients and becomes a potential tool to prevent doctors from false HT risk diagnosis.
Copyright © 2022 Elsevier Ltd. All rights reserved.

Author