The application of deep learning in the medical field has continuously made huge breakthroughs in recent years. Based on convolutional neural network (CNN), the U-Net framework has become the benchmark of the medical image segmentation task. However, this framework cannot fully learn global information and remote semantic information. The transformer structure has been demonstrated to capture global information relatively better than the U-Net, but the ability to learn local information is not as good as CNN. Therefore, we propose a novel network referred to as the O-Net, which combines the advantages of CNN and transformer to fully use both the global and the local information for improving medical image segmentation and classification. In the encoder part of our proposed O-Net framework, we combine the CNN and the Swin Transformer to acquire both global and local contextual features. In the decoder part, the results of the Swin Transformer and the CNN blocks are fused to get the final results. We have evaluated the proposed network on the synapse multi-organ CT dataset and the ISIC 2017 challenge dataset for the segmentation task. The classification network is simultaneously trained by using the encoder weights of the segmentation network. The experimental results show that our proposed O-Net achieves superior segmentation performance than state-of-the-art approaches, and the segmentation results are beneficial for improving the accuracy of the classification task. The codes and models of this study are available at https://github.com/ortonwang/O-Net.

Hematoxylin and eosin (H&E) stained colors is a critical step in the digitized pathological diagnosis of cancer. However, differences in section preparations, staining protocols and scanner specifications may result in the variations of stain colors in pathological images, which can potentially hamper the effectiveness of pathologist's diagnosis and the robustness. To alleviate this problem, several color normalization methods have been proposed. Most previous approaches map color information between images highly dependent on a reference template. However, due to the problem that pathological images are usually unpaired, these methods cannot produce satisfactory results. In this work, we propose an unsupervised color normalization method based on channel attention and long-range residual, using a technology called invertible neural networks (INN) to transfer the stain style while preserving the tissue semantics between different hospitals or centers, resulting in a virtual stained sample in the sense that no actual stains are used. In our method, the expert does not need to choose a template image. More specifically, we have developed a new unsupervised stain style transfer framework based on INN that is different from state-of-the-art methods. Meanwhile, we propose a new generator and a discriminator to further improve the performance. Our approach outperforms state-of-the-art methods both in objective metrics and subjective evaluations, yielding an improvement of 1.0 dB in terms of PSNR. Moreover, the amount of computation of the proposed network has been reduced by 33 %. This indicates that the inference speed is almost one third faster while the performance is better. INDEX TERMSColor normalization, stain style transfer, invertible neural networks, pathological images.

Early accurate diagnosis of patellofemoral pain syndrome (PFPS) is important to prevent the further development of the disease. However, traditional diagnostic methods for PFPS mostly rely on the subjective experience of doctors and subjective feelings of the patient, which do not have an accurate-unified standard, and the clinical accuracy is not high. With the development of artificial intelligence technology, artificial neural networks are increasingly applied in medical treatment to assist doctors in diagnosis, but selecting a suitable neural network model must be considered. In this paper, an intelligent diagnostic method for PFPS was proposed on the basis of a one-dimensional convolutional neural network (1D CNN), which used surface electromyography (sEMG) signals and lower limb joint angles as inputs, and discussed the model from three aspects, namely, accuracy, interpretability, and practicability. This article utilized the running and walking data of 41 subjects at their selected speed, including 26 PFPS patients (16 females and 10 males) and 16 painless controls (8 females and 7 males). In the proposed method, the knee flexion angle, hip flexion angle, ankle dorsiflexion angle, and sEMG signals of the seven muscles around the knee of three different data sets (walking data set, running data set, and walking and running mixed data set) were used as input of the 1D CNN. Focal loss function was introduced to the network to solve the problem of imbalance between positive and negative samples in the data set and make the network focus on learning the difficult-to-predict samples. Meanwhile, the attention mechanism was added to the network to observe the dimension feature that the network pays more attention to, thereby increasing the interpretability of the model. Finally, the depth features extracted by 1D CNN were combined with the traditional gender features to improve the accuracy of the model. After verification, the 1D CNN had the best performance on the running data set (accuracy = 92.4%, sensitivity = 97%, specificity = 84%). Compared with other methods, this method could provide new ideas for the development of models that assisted doctors in diagnosing PFPS without using complex biomechanical modeling and with high objective accuracy.

No abstract

No abstract