Lung cancer is one of the leading causes of global mortality, and accurate lung nodule segmentation in thoracic medical images is crucial for lung cancer analysis and diagnosis. In recent years, Vision Transformer (ViT) have been increasingly applied to medical image segmentation, demonstrating advantages in accuracy and scalability. However, existing ViT-based models still face limitations in segmenting heterogeneous targets with significant variations in shape and size. To address this, we propose a novel Transformer module, T-Unet, which leverages deformable convolutions. Our innovations include: (1) A deformable convolution-based attention mechanism that uses multi-scale feature fusion to generate offsets and improve attention precision; (2) Unique deformable patch embedding and position embedding layers designed for the Transformer module to enhance feature extraction; (3) Integration of this module into the UNet architecture to form T-Unet. We conducted experiments on the LIDC-IDRI, NSCLC and LUNA16 datasets. The results indicate that T-Unet performs remarkably in the lung nodule segmentation task. On the LIDC-IDRI dataset, it achieved a Dice coefficient of 91.89% and an IoU of 83.57%. On the NSCLC dataset, it achieved a Dice coefficient of 92.26% and an IoU of 82.73%. Additionally, on the LUNA16 dataset, T-Unet achieved a Dice coefficient of 92.08% and an IoU of 82.53%.

The widespread spread of disinformation has a significant negative impact on individuals and even society as a whole. Therefore, this paper proposes a novel hybrid model that organically combines blockchain technology with machine learning models. The hybrid model adopts a decentralized blockchain framework and smart contracts prior to information release to ensure its independence, and uses the incentive characteristics and trust scores of the blockchain to ensure that auditors send the correct audit results, and then uses machine learning technology to detect error information. We selected two public datasets, MediaEval-2015 and LIAR. Through experimental comparative analysis, the classification model with high accuracy was selected to effectively prevent the release of false information.

Recent studies have shown the superiority of neural networks on imaging equality and efficiency in inverse synthetic aperture radar (ISAR) resolution enhancement, but a central problem remains largely unsolved: all recent studies based on neural networks focused on minimizing the mean-squared reconstruction error (MSE), causing limited enhancing factors and inaccurate recovery of weak point scatters. In order to address this problem, a framework based on a generative adversarial network (GAN) using a combined loss composed of the absolute loss and the adversarial loss is proposed in this letter. The absolute loss ensures that reconstructed high-resolution ISAR images achieve higher enhancing factors and lower sidelobes. The adversarial loss pushes this framework to recover accurate amplitude and position of weak point scatters by a discriminator that is trained to differentiate reconstructed high-resolution ISAR images and real high-resolution ISAR images. Compared to some state-of-the-art methods, our GAN-based framework provides superior reconstruction with higher enhancing factors and more target details.