Single-source domain generalization aims to enhance model performance on unseen target domain test sets using only a single source domain dataset, typically by mitigating domain shifts between domains. In retinal vessel segmentation tasks, differences in dataset composition, such as variations in the proportions of different diseases and imaging noise levels, are considered significant sources of domain shift. However, few previous studies have delved into the mechanisms through which this type of domain shift influences model performance. In this study, we hypothesize that disparities in dataset composition could manifest as differences in distribution patterns of frequency domain features, rendering the model susceptible to overfitting specific patterns. Building on this hypothesis, we propose a novel Frequency Dropout based Single Source Domain Generalization (FD-SDG) framework that employs a Frequency Dropout Randomization mechanism to disentangle complex co-adaptive relationships among features from different frequency bands, thereby enhancing the model's robustness to variable frequency domain noise patterns in the sample space. Additionally, we introduce a Salient Structure Representation Normalization mechanism to align post-perturbation data features in the feature space using invariant anatomical structures. Through comparison experiments and ablation studies conducted on multiple sets of fundus images across-domain experiments, our method achieves state-of-the-art performance, underscoring its high generalizability and robustness.

To investigate the corneal morphology of adolescent diving athletes and analyze the related influencing factors. Corneal topographic maps were taken of 42 young athletes (19 males and 23 females, aged 9–17 years) from the diving team of Shenzhen Sports School, and then three morphological data of corneal curvature, astigmatism, and thickness were measured based on the topographic maps. A study was conducted to analyze the influence of diving on corneal morphology by comparing with the reference data of children of the same age in the literature, and statistically analyzing the relationship between different genders, ages, exercise levels, training time, and corneal morphology. The radius of the corneal curvature of athletes was significantly smaller than that of children of the same age (P < 0.01). There were statistical differences in the corneal curvatures of athletes in different age groups and training time groups (P < 0.05), and there were also differences in corneal thickness in different training time groups (P < 0.05). There were no statistically significant differences in corneal morphology among gender and exercise level groups. Diving can affect the corneal morphology of adolescent athletes, and the radius of corneal curvature is significantly lower than that of children of the same age. Corneal curvature is related to athletes' age and training time, and corneal thickness is affected by training time. Attention should be paid to the changes in the corneal morphology of diving athletes.

Deep learning models often encounter challenges in making accurate inferences when there are domain shifts between the source and target data. This issue is particularly pronounced in clinical settings due to the scarcity of annotated data resulting from the professional and private nature of medical data. Although various cross-domain strategies have been explored, including frequency-based approaches that vary appearance while preserving semantics, many remain limited by data constraints and computational cost. To tackle domain shifts in data-scarce medical scenarios, we propose a Random frequency filtering enabled Single-source Domain Generalization algorithm (RaffeSDG), which promises robust out-of-domain inference with segmentation models trained on a single-source domain. A frequency filter-based data augmentation strategy is first proposed to promote domain variability within a single-source domain by introducing variations in frequency space and blending homologous samples. Then Gaussian filter-based structural saliency is also leveraged to learn robust representations across augmented samples, further facilitating the training of generalizable segmentation models. To validate the effectiveness of RaffeSDG, we conducted extensive experiments involving out-of-domain inference on segmentation tasks for three human tissues imaged by four diverse modalities. Through thorough investigations and comparisons, compelling evidence was observed in these experiments, demonstrating the potential and generalizability of RaffeSDG. The code is available at https://github.com/liamheng/Non-IID_Medical_Image_Segmentation.

The annotation scarcity of medical image segmentation poses challenges in collecting sufficient training data for deep learning models. Specifically, models trained on limited data may not generalize well to other unseen data domains, resulting in a domain shift issue. Consequently, domain generalization (DG) is developed to boost the performance of segmentation models on unseen domains. However, the DG setup requires multiple source domains, which impedes the efficient deployment of segmentation algorithms in clinical scenarios. To address this challenge and improve the segmentation model's generalizability, we propose a novel approach called the Frequency-mixed Single-source Domain Generalization method (FreeSDG). By analyzing the frequency's effect on domain discrepancy, FreeSDG leverages a mixed frequency spectrum to augment the single-source domain. Additionally, self-supervision is constructed in the domain augmentation to learn robust context-aware representations for the segmentation task. Experimental results on five datasets of three modalities demonstrate the effectiveness of the proposed algorithm. FreeSDG outperforms state-of-the-art methods and significantly improves the segmentation model's generalizability. Therefore, FreeSDG provides a promising solution for enhancing the generalization of medical image segmentation models, especially when annotated data is scarce. The code is available at https://github.com/liamheng/Non-IID_Medical_Image_Segmentation.

As an important task in the field of remote sensing (RS) image processing, RS image change detection (CD) has made significant advances through the use of convolutional neural networks (CNNs). The transformer has recently been introduced into the field of CD due to its excellent global perception capabilities. Some works have attempted to combine CNN and transformer to jointly harvest local-global features; however, these works have not paid much attention to the interaction between the features extracted by both. Also, the use of the transformer has resulted in significant resource consumption. In this article, we propose the Asymmetric Cross-attention Hierarchical Network (ACAHNet) by combining CNN and transformer in a series-parallel manner. The proposed Asymmetric Multiheaded Cross Attention (AMCA) module reduces the quadratic computational complexity of the transformer to linear, and the module enhances the interaction between features extracted from the CNN and the transformer. Different from the early and late fusion strategies employed in previous work, the effectiveness of the mid-term fusion strategy employed by ACAHNet shows a new choice of timing for feature fusion in the CD task. Our experiments on the proposed method on three public datasets show that our network has a better performance in terms of effectiveness and computational resource consumption compared to other comparative methods.

Background: Ultrasound is widely used for image-guided therapy (IGT) in many surgical fields, thanks to its various advantages, such as portability, lack of radiation and real-time imaging. This article presents the first attempt to utilize multiple deep learning algorithms in distal humeral cartilage segmentation for dynamic, volumetric ultrasound images employed in minimally invasive surgery. Methods: The dataset, consisting 5,321 ultrasound images were collected from 12 healthy volunteers. These images were randomly split into training and validation sets in an 8:2 ratio. Based on deep learning algorithms, 9 semantic segmentation networks were developed and trained using our dataset at Southern University of Science and Technology Hospital in September 2022. The performance of the networks was evaluated based on their segmenting accuracy and processing efficiency. Furthermore, these networks were implemented in an IGT system to assess their feasibility in 3-dimentional imaging precision. Results: In 2D segmentation, Medical Transformer (MedT) showed the highest accuracy result with a Dice score of 89.4%, however, the efficiency in processing images was relatively lower at 2.6 frames per second (FPS). In 3D imaging, the average root mean square (RMS) between ultrasound (US)-generated models based on the networks and magnetic resonance imaging (MRI)-generated models was no more than 1.12 mm. Conclusions: The findings of this study indicate the technological feasibility of a novel method for real-time visualization of distal humeral cartilage. The increased precision of ultrasound calibration and segmentation are both important approaches to improve the accuracy of 3D imaging.

The morphology of the retinal vascular structure in fundus images is of great importance for ocular disease diagnosis. However, due to the poor fundus image quality and domain shifts between datasets, retinal vessel segmentation has long been regarded as a problematic machine-learning task. This work proposes a novel algorithm High-frequency Guided Cascaded Network (HGC-Net) to address the above issues. In our algorithm, a self-supervision mechanism is designed to improve the generalizability and robustness of the model. We apply Fourier Augmented Co-Teacher (FACT) augmentation to convert the style of fundus images, and extract high-frequency component (HFC) to highlight the vascular structure. The main structure of the algorithm is two cascaded U-nets, in which the first U-net generates a domain-invariant high-frequency map of fundus images, thus improving the segmentation stability of the second U-net. Comparison with the state-of-the-art methods and ablation study are conducted to demonstrate the excellent performance of our proposed HGC-Net.

Change detection (CD) is a particularly important task in the field of remote sensing image processing. It is of practical importance for people when making decisions about transitional situations on the Earth’s surface. The existing CD methods focus on the design of feature extraction network, ignoring the strategy fusion and attention enhancement of the extracted features, which will lead to the problems of incomplete boundary of changed area and missing detection of small targets in the final output change map. To overcome the above problems, we proposed a hierarchical attention residual nested U-Net (HARNU-Net) for remote sensing image CD. First, the backbone network is composed of a Siamese network and nested U-Net. We remold the convolution block in nested U-Net and proposed ACON-Relu residual convolution block (A-R), which reduces the missed detection rate of the backbone network in small change areas. Second, this paper proposed the adjacent feature fusion module (AFFM). Based on the adjacency fusion strategy, the module effectively integrates the details and semantic information of multi-level features, so as to realize the feature complementarity and spatial mutual enhancement between adjacent features. Finally, the hierarchical attention residual module (HARM) is proposed, which locally filters and enhances the features in a more fine-grained space to output a much better change map. Adequate experiments on three challenging benchmark public datasets, CDD, LEVIR-CD and BCDD, show that our method outperforms several other state-of-the-art methods and performs excellent in F1, IOU and visual image quality.

As an important task in the field of remote sensing image interpretation, change detection (CD) has been extensively studied by scholars in recent years. Affected by the illumination and the environment during bitemporal images' acquisition, there will be many pseudochanges, and the pseudochanges will seriously affect the effect of CD. Based on this, we propose a CD model named HMCNet, which introduces multilayer perceptron (MLP) into a convolutional neural network (CNN)-based CD model to form an MLP-CNN hybrid model. HMCNet has both the good feature extraction of CNN and the long-term dependence modeling ability of MLP, which can effectively overcome the interference of pseudochanges. In addition, the proposed cross-axis attention MLP can induce window attention of local features through shifted windows and, at the same time, form global attention to features through the interaction between information flows on the cross-axis, which effectively improves the comprehensive performance of MLP block. Extensive experiments on three public benchmark datasets show that HMCNet can achieve better performance with fewer parameters and Flops, and still maintain good generalization ability with fewer train data.

The morphological structure of retinal fundus blood vessels is of great significance for medical diagnosis, thus the automatic retinal vessel segmentation algorithm has become one of the research hotspots in the field of medical image processing. However, there are still several unsolved difficulties in this task: the existed methods are too sensitive to the low-frequency noise in the fundus images, and there are few annotated data sets available, and meanwhile, the retinal images of different datasets vary greatly. To solve the above problems, we propose a domain adaptive vessel segmentation algorithm with multiple image entrances called MIUnet, which is robust to the etiological noises and domain shift between diverse datasets. We apply Fourier domain adaptation and the high-frequency component filtering modules to transform the raw images into two styles, and simultaneously reduce the discrepancy between the source domain and target domain retinal images. After that, images produced by the two modules are fed into a multi-input deep segmentation model, and the full utilization of features from different modalities is ensured by the deep supervision mechanism. Experiments prove that, compared with other segmentation methods, the MIUnet has better performances in cross-domain experiments, where the IoU reaches 63% when trained on ARIA dataset and tested on the DRIVE dataset and 53% in the opposite direction.

Quality degradation (QD) is common in the fundus images collected from the clinical environment. Although diagnosis models based on convolutional neural networks (CNN) have been extensively used to interpret retinal fundus images, their performances under QD have not been assessed. To understand the effects of QD on the performance of CNN-based diagnosis model, a systematical study is proposed in this paper. In our study, the QD of fundus images is controlled by independently or simultaneously importing quantified interferences (e.g., image blurring, retinal artifacts, and light transmission disturbance). And the effects of diabetic retinopathy (DR) grading systems are thus analyzed according to the diagnosis performances on the degraded images. With images degraded by quantified interferences, several CNN-based DR grading models (e.g., AlexNet, SqueezeNet, VGG, DenseNet, and ResNet) are evaluated. The experiments demonstrate that image blurring causes a significant decrease in performance, while the impacts from light transmission disturbance and retinal artifacts are relatively slight. Superior performances are achieved by VGG, DenseNet, and ResNet in the absence of image degradation, and their robustness is presented under the controlled degradation.