Hemorrhagic transformation (HT) is a time-sensitive severe complication of endovascular thrombectomy for patients with ischemic stroke, and there is an urgent need to develop deep learning models to assist doctors in making rapid preliminary diagnoses. Currently popular Transformer deep learning architectures, while superior in modeling global relationships compared to traditional CNN, it still faces quadratic complexity issues when handling long sequences of medical images due to its inherent attention mechanism. In contrast, computational complexity of the Mamba model-based method grows linearly. Based on these findings, we have developed a novel Mamba model that effectively captures long-range dependencies and the sequential relationships among slices in high-dimensional medical image sequences. We evaluated the proposed model on a multi-center dataset. Experimental results show that our method outperforms other classical architectures and current advanced methods, validating the effectiveness and generalizability of the model composed of the aforementioned modules.

Hemorrhagic transformation (HT) is a severe complication of acute ischemic stroke (AIS) that can lead to disability or death. Accurate and timely risk assessment of HT is essential for clinicians to design effective treatment strategies. Previous studies in HT prediction have largely relied on machine learning and radiomics, which demand extensive manual data preprocessing by physicians. While some HT prediction models based on convolutional neural network (CNN) have been developed, they are limited in their ability to capture the sequential relationships between image slices and often lack the focus on crucial information. This study collected non-contrast computed tomography (NCCT) images and clinical data from 512 AIS patients across six hospitals to create a multi-center dataset. Based on the dataset, we propose CM-HTNet, a novel deep learning framework designed to predict HT risk in AIS patients following intravenous thrombolysis (IVT). CM-HTNet mimics the clinical process of reviewing NCCT images by focusing on key slices and integrating information from adjacent slices. It leverages CNNs to extract features from each NCCT slice and utilizes the Selective State-Space Model (SSM) within the Mamba framework to model sequential relationships between slices while prioritizing features relevant to HT prediction. Additionally, CM-HTNet incorporates the Neighborhood Rough Set (KRS) algorithm for clinical feature selection and cross-attention mechanisms to integrate clinical and imaging data for multimodal HT risk prediction. In testing on an external dataset from independent centers, CM-HTNet achieved a prediction accuracy of 88.85% and an AUC of 95.17%, showcasing its strong performance and generalization capabilities.

For the past decade, the K Nearest Neighbor (K-NN) search in high dimensional space has been explored extensively. Considerable theoretical and practical algorithms to accelerate approximate K-NN search have been presented. The approximate methods can improve searching efficiency and achieve satisfactory performance. Nevertheless, they are inherently approximation approaches and are not guaranteed to yield exact solutions. To obtain the K-NN results over high dimensional datasets efficiently while guaranteeing the same results as the linear search is a challenging task, attracting a large number of scholars. To this end, in this paper, we focus on improving the exact K-NN search efficiency over high dimensional datasets, and present a framework named LBPC to solve this problem. The lower bound based method and parallel computing are combined in LBPC to accelerate the exact K-NN search. In LBPC, the whole K-NN search task is divided into some sub-tasks and these sub-tasks can be conducted concurrently using the lower bound based method. The LBPC scheme allows users to utilize any lower bound to accelerate the exact K-NN query. In this paper, we use the segment mean to construct the lower bound and provide the theoretical analysis to show its computational efficiency and lower bound property. Various experiments are conducted to analyze the efficiency of LBPC, and the experimental results validate its effectiveness.

Structural magnetic resonance imaging (sMRI)-based Alzheimer's disease (AD) classification and its prodromal stage-mild cognitive impairment (MCI) classification have attracted many attentions and been widely investigated in recent years. Owing to the high dimensionality, representation of the sMRI image becomes a difficult issue in AD classification. Furthermore, regions of interest (ROI) reflected in the sMRI image are not characterized properly by spatial analysis techniques, which has been a main cause of weakening the discriminating ability of the extracted spatial feature. In this study, we propose a ROI-based contourlet subband energy (ROICSE) feature to represent the sMRI image in the frequency domain for AD classification. Specifically, a preprocessed sMRI image is first segmented into 90 ROIs by a constructed brain mask. Instead of extracting features from the 90 ROIs in the spatial domain, the contourlet transform is performed on each of these ROIs to obtain their energy subbands. And then for an ROI, a subband energy (SE) feature vector is constructed to capture its energy distribution and contour information. Afterwards, SE feature vectors of the 90 ROIs are concatenated to form a ROICSE feature of the sMRI image. Finally, support vector machine (SVM) classifier is used to classify 880 subjects from ADNI and OASIS databases. Experimental results show that the ROICSE approach outperforms six other state-of-the-art methods, demonstrating that energy and contour information of the ROI are important to capture differences between the sMRI images of AD and HC subjects. Meanwhile, brain regions related to AD can also be found using the ROICSE feature, indicating that the ROICSE feature can be a promising assistant imaging marker for the AD diagnosis via the sMRI image. Code and Sample IDs of this paper can be downloaded at https://github.com/NWPU-903PR/ROICSE.git.

Local binary pattern (LBP) is a frequently‐used texture descriptor. Lots of LBP‐variants have been proposed to improve its performance of representing textures. However, most of them ignore the global and neighbour‐difference information of an image texture. In this study, the authors propose a structural difference histogram representation by fusing the segmented structure pattern (SSP), the refined LBP (RLBP) and the neighbour‐difference pattern (NDP) for texture classification. Particularly, the segmented structure, which contains the global contour information of an image texture, is first constructed to compute its SSP histogram feature. Simultaneously, the RLBP is defined to represent the local texture information. Furthermore, the NDP is presented to describe differences between neighbours of a centre pixel in the local patch of texture images. Experimental results on Brodatz and Columbia‐Utrecht reflectance databases indicate that the proposed method can achieve the satisfactory classification accuracy compared with several representative methods.

The widely used energy features extracted from the wavelet domain can effectively represent the common image textures. However, they are not robust to the rotated textures. In this letter, we propose a multiscale rotation-invariant representation (MRIR) of textures by using multiscale sampling. Particularly, a multiscale wavelet transform is used to decompose the magnitude pattern (MP) mapping of a texture. Furthermore, the sign pattern (SP) mapping of a texture is used as a step function, which is further sampled and used to fit the wavelet subbands of the MP mapping for computing the sampled directional mean vectors (SDMVs) of the subbands. Moreover, we construct frequency vectors (FVs) of those SP mappings for capturing the structural information of textures. Finally, we can obtain the MRIR vector of an image texture by concatenating those SDMVs and FVs for texture classification. The comprehensive experimental results demonstrate that our proposed approach outperforms six representative texture classification methods.