Ischemic stroke constitutes 62.4 % of all strokes and intracerebral hemorrhage 27.9 % [1,2], with radically different etiological, therapeutic management [3,4], and prognosis. Intracerebral hemorrhage is associated with higher morbidity and mortality than ischemic stroke [5], which itself has a poorer prognosis in the presence of hemorrhagic transformation. Importantly, the distinction between hemorrhagic and ischemic stroke is a keystone of treatment decision, since hemorrhage contraindicates thrombolysis [6]. Early detection of intracranial bleeding in the acute phase of stroke is therefore crucial.

MRI allows to distinguish hemorrhagic from ischemic stroke [6,7], with T2*-weighted images detecting acute hematomas through deoxyhemoglobin-mediated susceptibility effects. Although MRI acquisition times are longer compared to computed tomography (CT) [8], non-randomized studies suggest that it may result in better outcomes following mechanical thrombectomy [9].

In the context of acute stroke, “time is brain” [10] and all efforts to reduce time from onset to treatment decision are needed. This includes shortening the MRI acquisition protocol [11] while keeping the relevant information, including those needed to detect brain hemorrhage. To reduce the time devoted to MRI, multiple approaches have been proposed, including parallel imaging [12] and echo planar imaging (EPI) [13,14]. Instead of reducing the duration of each individual sequence, artificial intelligence (AI) offers new possibilities [15], such as generating MR images from CT images [16], or directly substituting them using generative algorithm. In a previous study, a generative fluid-attenuated with inversion recovery (FLAIR) sequence, based on diffusion-weighted (DW) imaging, was proposed, potentially eliminating the need for acquiring the true FLAIR sequence [17], thereby optimizing MRI protocol acquisition time [18,19].

Along the same principle, we hypothesized that generative T2*-weighted images created by an edge-aware generative adversarial network (Ea-GAN) [20] from DW images could enable the detection of brain hemorrhage. Indeed, b = 0 and b = 1000 s/mm2 images of the DW sequence are acquired using EPI, which is inherently sensitive to local magnetic field inhomogeneities [21] and therefore to hemoglobin degradation products, suggesting redundancy with T2*-weighted sequence [22].

The purpose of this study was to validate a deep-learning algorithm generating T2*-weighted images from DW images and compare its performance to true T2*-weighted images for hemorrhage detection in patients with acute stroke.