Recently, convolutional neural networks (CNNs) have achieved state-of-the-art performance in infrared small target detection. However, the limited number of public training data restricts the performance improvement of CNN-based methods. To handle the scarcity of training data, we propose a method that can generate synthetic training data for infrared small target detection. We adopt the generative adversarial network framework where synthetic background images and infrared small targets are generated in two independent processes. In the first stage, we synthesize infrared images by transforming visible images into infrared ones. In the second stage, target masks are implanted on the transformed images. Then, the proposed intensity modulation network synthesizes realistic target objects that can be diversely generated from further image processing. Experimental results on the recent public dataset show that, when we train various detection networks using the dataset composed of both real and synthetic images, detection networks yield better performance than using real data only.

Deep neural networks have shown promising results in the analysis of structural magnetic resonance imaging (sMRI) data for the diagnosis of dementia, particularly Alzheimer’s disease (AD). Different regions of the brain have diverse structures that are linked to specific functions, which could account for the variability in disease-related changes observed in sMRI scans of these areas. Understanding the overall characteristics of sMRI data is important since current popular convolutional neural networks (CNN) for deep learning do not consider the interconnection of voxels. Vision transformers have shown effectiveness in identifying long-distance connections in the brain, which has led to their success in applications such as disease detection. However, the image noise and limited scalability of self-attention mechanisms in relation to image size has hindered their widespread use in advanced Alzheimer’s analysis. To enhance information retention and reduce network complexity, this study presents a novel adaptable efficient receptive field feature extraction network. Moreover, an advanced attention mechanism with both local grid attention block and dilated global attention module has been incorporated to highlight the characteristics of AD. Next, a more improved hierarchical inverted residual feed forward network in place of multi-layer perceptron is suggested to enhance the characterization of features through the integration of information from both lower and higher layers. Finally, the global average pooling and $1\times 1$ convolution are used to reduce dimensionality, enhance non-linearity, and allow channel interactions in feature maps before being input into the classification head. The network achieved high performance in various scenarios, with average accuracies of 97.29% for AD vs. HC and 94.79% for MCI vs. HC classification using ADNI as experimental datasets.

Hippocampus segmentation on MRI (magnetic resonance imaging) plays a vital role in detecting, diagnosing, tracking, and monitoring neurodegenerative diseases, particularly Alzheimer’s disease. While larger datasets often provide an advantage in deep learning-based segmentation, smaller datasets pose unique challenges due to limited data variability and an increased risk of overfitting. This study addresses these challenges by developing a computationally efficient and accurate 3D U-Net model tailored for hippocampus segmentation. The proposed approach employs a preprocessing pipeline combining 3D Contrast Limited Adaptive Histogram Equalization (CLAHE) and Selective Coefficient-Enhanced 3D Wavelet Transform (SCE-3DWT), which enhances contrast and reduces noise for improved feature extraction. The experimental evaluation was conducted using the EADC-ADNI HarP dataset, comprising 135 hippocampal MRI scans with an input image size of 64×64×96. The model achieved a Dice coefficient of 0.8838 and a Jaccard Index of 0.7920, surpassing recent state-of-the-art methods. Comparative analysis highlights reduced Over-Segmentation Ratio (OSR = FP/(FP+TP), 0.0594) and Under-Segmentation Ratio (USR = FN/(FN+TP), 0.0569, reflecting its robustness and generalization. The lightweight architecture, designed with a maximum filter size of 512, operates efficiently without relying on transfer learning, making it accessible for broader applications. Future work will focus on integrating post processing techniques, leveraging larger and more diverse datasets, and exploring higher-resolution volumetric data to further improve segmentation accuracy and clinical utility. This study contributes to the advancement of medical image analysis, offering a resource-efficient framework for precise hippocampus segmentation, with potential implications for improved Alzheimer’s disease management.

In the various industrial manufacturing processes, the automatic visual inspection system is an essential part as it reduces the chances of delivery of defective products and the cost of training and hiring experts for manual inspection. In this work, we propose a new unsupervised anomaly detection method inspired by the masked language model for the automatic visual inspection system. The proposed method consists of an image tokenizer and two subnetworks, a reconstruction subnetwork, and a segmentation subnetwork. We adopt a pre-trained self-supervised vision Transformer model to use it as an image tokenizer. Our first subnetwork is trained to predict the anomaly-free patch tokens and the second subnetwork is trained to produce anomaly segmentation results from both the reconstructed and input patch tokens. During training, only the two subnetworks are optimized, and parameters of an image tokenizer are frozen. Experimental results show that the proposed method exhibits better performance than conventional methods in detecting defective products by achieving 99.05% I-AUROC on MVTecAD dataset and 94.8% I-AUROC on BTAD.

The basic concept of the scalable video coding (SVC) is to reduce the amount of bits for video representation by exploiting the correlation between coding layers. In the SVC, a macroblock (MB) of the enhancement layer (EL) is predicted using the prediction MB which is the upsampled version of the co-located MB of the base layer (BL). In this paper, we propose additional prediction MBs which are displaced from co-located position. In the proposed method, displacement vector (DV) indicating the best matching MB of the BL is encoded when another MB is more similar to the MB of the EL than the co- located one. The bits for DV are adaptively assigned based on the Lagrangian cost function. Experimental results show that the proposed algorithm achieves better performance in the rate- distortion sense.