A New Multi-modal Dataset and Two-stage DNN Approach for Acute Ischemic Stroke Detection

Multi-modal data play an essential role in medical diagnostics, in particular for the detection of acute ischemic stroke (AIS). However, existing methods for AIS detection focus on single-modality learning, neglecting the advantages of integrating multiple modalities as well as lacking multi-modal database. In this paper, we introduce a novel multi-modal dataset consisting of 80 cases with 5 medical modalities and propose a two-stage deep neural network (DNN) framework designed for AIS detection. The proposed framework includes two subnet: A lesion localization subnet for the preliminary identification of potential lesion regions, and a stroke segmentation subnet for the precise delineation of stroke areas. Additionally, we incorporate a Transformer-based multi-modal fusion module that effectively learns cross-attention between different modalities. The proposed method is evaluated on our newly established multi-modal dataset, and experimental results demonstrate that our method achieves the state-of-the-art performance for AIS detection.