The diagnosis of high myopia using fundus images is essential for visual health. Existing deep learning-based methods rely on large-scale labeled data, but annotated data for high myopia fundus images remains scarce. To address this issue, we mimic the ability of ophthalmologists to diagnose a new disease with only a small number of samples. In this regard, we propose a meta-learning network guided by domain knowledge from fundus images for diagnosing high myopia. The model consists of three modules: First, the image reconstruction module builds an autoencoder (comprising an encoder and a decoder) that takes the input raw image and outputs the reconstructed image. Next, the fundus image domain knowledge learning module constructs a Siamese network to learn the similarity between the original and reconstructed fundus images. This similarity is used as a loss function to guide the encoder in effectively learning the domain features of fundus images. Finally, in the domain-knowledge-guided meta-learning module, the encoder’s initialization parameters (obtained from the first two modules) are further optimized using the OCMAMAL architecture, resulting in more optimal encoder parameters. This optimization helps achieve superior recognition performance with only a small amount of data for new tasks. Using a clinically real high myopia fundus image dataset, our method achieved F1 scores of 83.4%, 87.9%, and 89.2% under 5-shot, 10-shot, and 20-shot conditions, respectively, demonstrating the effectiveness of the proposed method.

Diabetic retinopathy (DR) is the leading cause of blindness in diabetic patients, which necessitates precise grading of retinal lesions for early diagnosis. Existing DR grading methods typically employ image enhancement techniques to improve the quality of fundus images. However, due to variations in imaging devices and differences in the proficiency of medical practitioners, the quality of images often exhibits significant heterogeneity. Uniform enhancement across all fundus images may inadvertently amplify noise artifacts, particularly in high-quality images. Moreover, since diabetic lesions in fundus images are often small, reliance solely on global image features makes it difficult to fully capture fine-grained lesion features. To address these challenges, this paper proposes a dual-stream deep learning model that integrates quality-aware dynamic enhancement and a multi-instance multi-scale vision transformer. First, An image quality assessment-based selective enhancement strategy was implemented, wherein only low-quality fundus images underwent enhancement processing. Then, a dual-branch processing architecture is designed to differentially handle enhanced and non-enhanced images. Experimental results on real-world datasets demonstrate the effectiveness of the proposed method.

Non-invasive fundus images can be used to diagnose various fundus diseases, such as high myopia (HM). Existing deep learning-based research mainly relies on data to drive the model to learn key features. However, the data related to HM is limited (especially for young children), making it difficult for deep networks to accurately focus on key features. Hence, we propose a prior knowledge-guided deep learning network for pediatric HM detection. It comprises four modules: (1) Prior Feature-Based Channel Fusion: This module extracts key features (brightness, edges, texture) from fundus images using image processing methods to obtain corresponding single-channel slices. Through channel-level feature fusion, these slices are used to construct multiple sets of feature-enhanced datasets. (2) Global Fundus Feature Extraction: It uses residual blocks to build the backbone, and builds aU-shaped attention component based on the U -shaped network. This module extracts the global and context information of the original fundus image to obtain a global feature map. (3) Knowledge-Guided Attention Generation: The residual structure is employed to further extract the hidden features of the feature-enhanced data, thereby obtaining local key feature maps. (4) Pediatric HM Classification: By combining local key feature maps (obtained in module 3) with global feature maps (obtained in module 2) through spatial attention mechanism, the deep network is guided to complete the classification task of pediatric HM. Extensive experiments on real-world datasets demonstrate the effectiveness of our method (accuracy is 0.921, F1 score is 0.903).

Multi-class segmentation of eye images plays a pivotal role in assessing patients with myasthenia gravis, and the measurement results rely heavily on the segmentation accuracy. However, there is still a problem with inaccurate boundary segmentation. Compared to heuristic-based network structure optimization, exploring effective loss function is an intuitive, simple, and interpretable way to address this issue. In this paper, we experimentally verify the effectiveness of boundary loss for multi-class segmentation of eye images and investigate its hybrid law with other segmentation losses. The application of the study significantly enhances the accuracy of myasthenia gravis scoring and holds promise for assisting in the evaluation of various other eye diseases.

Sign language video translation, which converts sign language information into textual expressions, play a vital role in breaking down the language communication barrier between deaf and healthy people. Existing translation methods are mainly focus on the single and pure background. However, the background in real-world environments is always complex, and these methods are difficult to achieve effective recognition results. To address this issue, we have exploratively constructed a real-world complex background sign language dataset (CBSL), containing sign language videos captured in various authentic environments (e.g., different backgrounds and lighting conditions). Based on this, we propose a progressive sign language translation model to effectively separate sign language users from the background and reduce environmental interference, thus significantly improving the generalization ability. Our proposed method significantly outperforms various comparative methods across all performance metrics on the CBSL dataset. Furthermore, on the publicly available Chinese Sign Language Continuous Recognition dataset(CSL), our method performs comparably to the current state-of-the-art (SOTA).

Microscopic imaging plays an important role in the biomedical field. Existing deep learning based methods rely on high-quality data. However, there is a lot of noise (such as bubbles and impurities) in the microscopic images of biological samples collected outdoors, which may lead to significant interference in the microscopic objects identification task. To solve this problem, this paper proposes a deep learning based method for microscopic object localization and classification. Firstly, the whole slide image is preprocessed to obtain the microscopic images after preliminary filtering bubbles and impurities. Then, the sensitized pollen grains are located based on the deep learning model to remove the interference of remaining impurities, and the microscopic images of sensitized pollen grains are classified. This method can effectively suppress the interference of noise in microscopic images on object classification and improve the accuracy and reliability of model. The proposed method is verified by experiments based on real data and the results show that the proposed method achieves the highest accuracy compared with other deep learning methods.

Pollen classification plays an essential role in many fields such as medicine and palynology. Notably, manual pollen identification via observing key pollen information (e.g., their contours) is time-consuming and laborious. To date, deep learning methods can extract complex features in an end-to-end manner. However, deep learning based automatic classification methods on pollen grains are still rare, and their performances remain unsatisfactory owing to limitation of interference from irrelevant information (such as impurities and bubbles) and the lack of pollen attention. Based on the above considerations, we propose a contour-guided network called CG-Net, which contains three modules. Image pre-processing module first removes impurities and bubbles in pollen images according to the color information. Then, contour awareness module is designed to generate contour features and these features are served as attention maps for next module. Finally, contour guidance module weights the yielded contour attention maps to both original images and feature maps of different convolution layers, making the CNN focus on discriminative features of pollen grains. Extensive experiments are conducted on several real-world pollen datasets, and the results demonstrate the effectiveness of our proposed method with the accuracy and F1-score over 84%.