Ensemble learning for retinal disease recognition under limited resources.

Retinal optical coherence tomography (OCT) images provide crucial insights into the health of the posterior ocular segment. Therefore, the advancement of automated image analysis methods is imperative to equip clinicians and researchers with quantitative data, thereby facilitating informed decision-making. The application of deep learning (DL)-based approaches has gained extensive traction for executing these analysis tasks, demonstrating remarkable performance compared to labor-intensive manual analyses. However, the acquisition of retinal OCT images often presents challenges stemming from privacy concerns and the resource-intensive labeling procedures, which contradicts the prevailing notion that DL models necessitate substantial data volumes for achieving superior performance. Moreover, limitations in available computational resources constrain the progress of high-performance medical artificial intelligence, particularly in less developed regions and countries. This paper introduces a novel ensemble learning mechanism designed for recognizing retinal diseases under limited resources (e.g., data, computation). The mechanism leverages insights from multiple pre-trained models, facilitating the transfer and adaptation of their knowledge to retinal OCT images. This approach establishes a robust model even when confronted with limited labeled data, eliminating the need for an extensive array of parameters, as required in learning from scratch. Comprehensive experimentation on real-world datasets demonstrates that the ensemble models constructed by the proposed ensemble method show superior performance over the baseline models under sparse labeled data, especially the triple ensemble model, which achieves the accuracy of 92.06%, which is 8.27%, 7.99%, and 11.14% better than the three baseline models, respectively. In addition, compared with the three baseline models learned from scratch, the triple ensemble model has fewer trainable parameters, only 3.677M, which is lower than the three baseline models of 8.013M, 4.302M, and 20.158M, respectively.

Interpretable CNN-Multilevel Attention Transformer for Rapid Recognition of Pneumonia From Chest X-Ray Images.

Chest imaging plays an essential role in diagnosing and predicting patients with COVID-19 with evidence of worsening respiratory status. Many deep learning-based approaches for pneumonia recognition have been developed to enable computer-aided diagnosis. However, the long training and inference time makes them inflexible, and the lack of interpretability reduces their credibility in clinical medical practice. This paper aims to develop a pneumonia recognition framework with interpretability, which can understand the complex relationship between lung features and related diseases in chest X-ray (CXR) images to provide high-speed analytics support for medical practice. To reduce the computational complexity to accelerate the recognition process, a novel multi-level self-attention mechanism within Transformer has been proposed to accelerate convergence and emphasize the task-related feature regions. Moreover, a practical CXR image data augmentation has been adopted to address the scarcity of medical image data problems to boost the model's performance. The effectiveness of the proposed method has been demonstrated on the classic COVID-19 recognition task using the widespread pneumonia CXR image dataset. In addition, abundant ablation experiments validate the effectiveness and necessity of all of the components of the proposed method.

High-efficiency FBG array sensor interrogation system via a neural network working with sparse data.

FBG array sensors have been widely used in the multi-point monitoring of large structures due to their excellent optical multiplexing capability. This paper proposes a cost-effective demodulation system for FBG array sensors based on a Neural Network (NN). The stress variations applied to the FBG array sensor are encoded by the array waveguide grating (AWG) as transmitted intensities under different channels and fed to an end-to-end NN model, which receives them and simultaneously establishes a complex nonlinear relationship between the transmitted intensity and the actual wavelength to achieve absolute interrogation of the peak wavelength. In addition, a low-cost data augmentation strategy is introduced to break the data size bottleneck common in data-driven methods so that the NN can still achieve superior performance with small-scale data. In summary, the demodulation system provides an efficient and reliable solution for multi-point monitoring of large structures based on FBG array sensors.

Fabry-Perot interferometric sensor demodulation system utilizing multi-peak wavelength tracking and neural network algorithm.

For FPI sensor demodulation systems to be used in actual engineering measurement, they must have high performance, low cost, stability, and scalability. Excellent performance, however, necessitates expensive equipment and advanced algorithms. This research provides a new absolute demodulation system for FPI sensors that is high-performance and cost-effective. The reflected light from the sensor was demultiplexed into distinct channels using an array waveguide grating (AWG), with the interference spectrum features change translated as the variation of the transmitted intensity in each AWG channel. This data was fed into an end-to-end neural network model, which was utilized to interrogate multiple interference peaks' absolute peak wavelengths simultaneously. This architecturally simple network model can achieve remarkable generalization capabilities without training large-scale datasets using an appropriate data augmentation strategy. Experiments show that in simultaneous multi-wavelength and cavity length interrogations, the proposed system has the precision of up to ± 14 pm and ± 0.07 µm, respectively. The interrogation resolution can theoretically reach the pm level benefit from the neural network method. Furthermore, the system's outstanding demodulation repeatability and suitability were demonstrated. The system is expected to provide a high-performance and cost-effective, reliable solution for practical engineering applications.

Efficient calculation of optical properties of suspended-core fiber via a machine learning algorithm.

Growing nonlinearity demands in mid-infrared applications place more outstanding requirements on fiber structure design. Chalcogenide suspended-core fibers (SCFs) are considered excellent candidates for mid-infrared applications due to their significant advantages in nonlinearity and dispersion management. However, traditional numerical methods for accurate modeling and optimization of SCFs often rely on the performance of computing devices and have many limitations when dealing with complex models. A machine learning algorithm is applied to calculate the optical properties of chalcogenide SCFs, including effective mode area, nonlinear coefficient, and dispersion. The established artificial neural network (ANN) model enables accurate prediction of the above optical properties of As2S3 SCF, for which in the wavelength range of 1.0 to 4.0 µm, the radius of the fiber core is 0.4 to 0.6 µm, and width of the cantilever is 0.06 to 0.09 µm. We demonstrate that this simple ANN model has considerable advantages over the traditional numerical calculation model in computational speed and resource utilization. In summary, the proposed model can quickly provide more accurate optical property predictions, providing a cost-effective solution for precise modeling and optimization of chalcogenide SCFs.

Domain-Adaptive Prototype-Recalibrated Network with Transductive Learning Paradigm for Intelligent Fault Diagnosis under Various Limited Data Conditions.

In real industrial scenarios, intelligent fault diagnosis based on data-driven methods has been widely researched in the past decade. However, data scarcity is widespread in fault diagnosis tasks owning to the difficulties in collecting adequate data. As a result, there is an increasing demand for both researchers and engineers for fault identification with scarce data. To address this issue, an innovative domain-adaptive prototype-recalibrated network (DAPRN) based on a transductive learning paradigm and prototype recalibration strategy (PRS) is proposed, which has the potential to promote the generalization ability from the source domain to target domain in a few-shot fault diagnosis. Within this scheme, the DAPRN is composed of a feature extractor, a domain discriminator, and a label predictor. Concretely, the feature extractor is jointly optimized by the minimization of few-shot classification loss and the maximization of domain-discriminative loss. The cosine similarity-based label predictor, which is promoted by the PRS, is exploited to avoid the bias of naïve prototypes in the metric space and recognize the health conditions of machinery in the meta-testing process. The efficacy and advantage of DAPRN are validated by extensive experiments on bearing and gearbox datasets compared with seven popular and well-established few-shot fault diagnosis methods. In practical application, the proposed DAPRN is expected to solve more challenging few-shot fault diagnosis scenarios and facilitate practical fault identification problems in modern manufacturing.

Semi-Supervised Deep Learning Model for Efficient Computation of Optical Properties of Suspended-Core Fibers.

Suspended-core fibers (SCFs) are considered the best candidates for enhancing fiber nonlinearity in mid-infrared applications. Accurate modeling and optimization of its structure is a key part of the SCF structure design process. Due to the drawbacks of traditional numerical simulation methods, such as low speed and large errors, the deep learning-based inverse design of SCFs has become mainstream. However, the advantage of deep learning models over traditional optimization methods relies heavily on large-scale a priori datasets to train the models, a common bottleneck of data-driven methods. This paper presents a comprehensive deep learning model for the efficient inverse design of SCFs. A semi-supervised learning strategy is introduced to alleviate the burden of data acquisition. Taking SCF's three key optical properties (effective mode area, nonlinear coefficient, and dispersion) as examples, we demonstrate that satisfactory computational results can be obtained based on small-scale training data. The proposed scheme can provide a new and effective platform for data-limited physical computing tasks.

Cost-effective improvement of the performance of AWG-based FBG wavelength interrogation via a cascaded neural network.

Fiber Bragg grating (FBG) sensors have been widely applied in various applications, especially for structural health monitoring. Low cost, wide range, and low error are necessary for an excellent performance FBG sensor signal demodulation system. Yet the improvement of performance is commonly accompanied by costly and complex systems. A high-performance, low-cost wavelength interrogation method for FBG sensors was introduced in this paper. The information from the FBG sensor signal was extracted by the array waveguide grating (AWG) and fed into the proposed cascaded neural network. The proposed network was constructed by cascading a convolutional neural network and a residual backpropagation neural network. We demonstrate that our network yields a vastly significant performance improvement in AWG-based wavelength interrogation over that given by other machine learning models and validate it in experiments. The proposed network cost-effectively widens the wavelength interrogation range of the demodulation system and optimizes the wavelength interrogation error substantially, also making the system scalable.