Super-resolution reconstruction of ultrasound images can effectively improve image quality by enhancing the high-frequency information and enriching the detailed features of the images. To address the problem that existing super-resolution reconstruction of ultrasound images is prone to detail distortion in this work, a method for super-resolution reconstruction of lung ultrasound images based on texture feature enhanced generative adversarial networks (TFEGAN) is proposed. This method employs a multi-scale texture feature extraction module that fully extracts texture information at multiple scales form lung ultrasound images using a multi-branch architecutre. It employs channel attention mechanisms and multi-head self-attention mechanisms to enhance the feature extraction of the super-resolution generative adversarial network. As a result the network is able to dynamically adjust the weights of different channel features and capture long-range dependencies, thereby enhancing its global feature representation capabilities. Finally, by combining a joint loss function with an adaptive loss-weight adjustment strategy, super-resolution reconstruction of lung ultrasound images was achieved. Experimental results show that compared to conventional algorithms such as super-resolution generative adversarial network (SRGAN), enhanced super-resolution generative adversarial networks (ESRGAN), structure-preserving super resolution (SPSR) and content-aware local GAN (CAL-GAN), the learned perceptual image patch similarity (LPIPS) index of our new method is improved by 17.3%, 3.76%, 9.70%, and 2.85%. The reconstructed image texture details are discernible, and the overall image quality is enhanced.

Thyroid nodule elastography ultrasound images contain information about nodule stiffness and morphology. Accurate segmentation of these images can significantly enhance the accuracy of thyroid nodule diagnosis. While deep learning-based thyroid nodule ultrasound image segmentation has shown promising results, its accuracy in elastography image segmentation remains limited due to the small dataset sizes. To address the low segmentation accuracy of thyroid nodule elastography images, we propose a transfer learning-based method that leverages shared features between grayscale ultrasound and elastography images. We first introduce a large-kernel attention mechanism to develop a multi-scale feature extraction network, capturing both local structural details and long-range dependencies in thyroid nodules. U-Net is then employed as the backbone network to build the grayscale ultrasound segmentation model. On this basis, we use the conditional generative adversarial network (CGAN) to align the feature distributions of grayscale ultrasound images (the source domain) and elastography ultrasound images (the target domain). A weight-sharing strategy is applied to transfer the feature extraction parameters, establishing a segmentation model for elastography ultrasound images. Experimental results show that the proposed segmentation method affords Dice coefficient, intersection over union (IOU), Recall, and Precision values of 77.09%, 65.20%, 78.19%, and 81.05%, respectively.

To address the problem of noise degrading fusion performance in medical images, we propose a multimodal medical image fusion method based on cross-modal channel-aware (CMCA) module. A dual-branch encoder enhanced with a squeeze-and-excitation module is introduced to apply channel-wise weighting during feature extraction and obtain modality-specific features. A cross-modal channel-aware fusion module is constructed to integrate the extracted information and achieve complementary feature fusion. A composite loss function combining image entropy and median-based weighting is adopted to preserve detail while suppressing noise during training, thereby enabling effective multimodal medical image fusion. Experimental results on an MRI-CT dataset show that the proposed method achieves an average gradient of 8.630, a standard deviation of 82.301 with a spatial frequency of 35.728, and a structural similarity index of 1.173, which makes it a valuable tool for assisting clinicians in lesion analysis.

Batch processes have multimode characteristics, and the existing methods of mode partition for batch processes ignore the causal relationship between the process characteristics and the mode centers, which directly affects the accuracy and reasonableness of the mode partition results and the prediction precision and generalizability of the models. A method for the online prediction of quality variables based on deep causal clustering-relevance vector machine (DCC-RVM) for multimode batch processes is proposed in this work. By combining density peak clustering, with the strong nonlinearity of the batch process data, the deep features of the process data are first extracted using a deep autoencoder. The process model for batch processes based on long short-term memory (LSTM) is then constructed, with the causal relationship between mode centers and the LSTM process model as a constraint, and the mode centers selection strategy is constructed to obtain reasonable mode centers; Subsequently, by considering the temporal features of the process data, the non-mode center samples are assigned to the corresponding modes based on the relative distance between samples, and the partition points and data sets are obtained for each mode; Finally, the quality variable prediction model for each mode is established using RVM, and the batch process quality variable is predicted online. The experimental results show that the mode partition by DCC is reasonable. The root mean square error (RMSE) and R² for predictiions using our DCC-RVM model are 0.0104 and 0.9995, respectively.