Popular Deep Learning Research Articles

Glaucoma detection and staging from visual field images using machine learning techniques.

In this study, we investigated the performance of deep learning (DL) models to differentiate between normal and glaucomatous visual fields (VFs) and classify glaucoma from early to the advanced stage to observe if the DL model can stage glaucoma as Mills criteria using only the pattern deviation (PD) plots. The DL model results were compared with a machine learning (ML) classifier trained on conventional VF parameters. A total of 265 PD plots and 265 numerical datasets of Humphrey 24-2 VF images were collected from 119 normal and 146 glaucomatous eyes to train the DL models to classify the images into four groups: normal, early glaucoma, moderate glaucoma, and advanced glaucoma. The two popular pre-trained DL models: ResNet18 and VGG16, were used to train the PD images using five-fold cross-validation (CV) and observed the performance using balanced, pre-augmented data (n = 476 images), imbalanced original data (n = 265) and feature extraction. The trained images were further investigated using the Grad-CAM visualization technique. Moreover, four ML models were trained from the global indices: mean deviation (MD), pattern standard deviation (PSD) and visual field index (VFI), using five-fold CV to compare the classification performance with the DL model's result. The DL model, ResNet18 trained from balanced, pre-augmented PD images, achieved high accuracy in classifying the groups with an overall F1-score: 96.8%, precision: 97.0%, recall: 96.9%, and specificity: 99.0%. The highest F1 score was 87.8% for ResNet18 with the original dataset and 88.7% for VGG16 with feature extraction. The DL models successfully localized the affected VF loss in PD plots. Among the ML models, the random forest (RF) classifier performed best with an F1 score of 96%. The DL model trained from PD plots was promising in differentiating normal and glaucomatous groups and performed similarly to conventional global indices. Hence, the evidence-based DL model trained from PD images demonstrated that the DL model could stage glaucoma using only PD plots like Mills criteria. This automated DL model will assist clinicians in precision glaucoma detection and progression management during extensive glaucoma screening.

Research on automatic modulation recognition of shortwave signals in few‐shot scenarios based on knowledge distillation

AbstractAutomatic modulation recognition plays an important role in wireless communication and radio regulation. Existing deep learning‐based automatic modulation recognition techniques perform well with large datasets and high computational power but require significant resources for labelled data and complex pre‐processing. This paper proposes a multi‐information fusion method for few‐shot modulation recognition, which involves converting I/Q signals into A/P signals and training with a combination of VGG and LSTM network models. An ensemble knowledge distillation (EKD) approach is employed to streamline the network model, meeting the demands for deploying neural network models on compact devices. Experimental results demonstrate that using only 1% of the shortwave modulation signal dataset as the training set, the proposed model achieves an average classification accuracy of 71.08% under all signal‐to‐noise ratios, surpassing the currently popular deep learning models. Moreover, two small‐scale networks, MobileNetV3 and convolutional neural network are trained, through EKD. Compared to the teacher network, the floating‐point operations of the distilled models are reduced by 99.8% and 99.7%, respectively, and the average prediction accuracy only decreases by 16.05% and 8.09%. The lightweight, few‐shot networks designed in this study for shortwave modulation signals aim to achieve fast and accurate modulation recognition on compact devices.

History-Driven Fuzzing For Deep Learning Libraries

Recently, many Deep Learning (DL) fuzzers have been proposed for API-level testing of DL libraries. However, they either perform unguided input generation (e.g., not considering the relationship between API arguments when generating inputs) or only support a limited set of corner-case test inputs. Furthermore, many developer APIs crucial for library development remain untested, as they are typically not well documented and lack clear usage guidelines, unlike end-user APIs. This makes them a more challenging target for automated testing. To fill this gap, we propose a novel fuzzer named Orion, which combines guided test input generation and corner-case test input generation based on a set of fuzzing heuristic rules constructed from historical data known to trigger critical issues in the underlying implementation of DL APIs. To extract the fuzzing heuristic rules, we first conduct an empirical study on the root cause analysis of 376 vulnerabilities in two of the most popular DL libraries, PyTorch and TensorFlow. We then construct the fuzzing heuristic rules based on the root causes of the extracted historical vulnerabilities. Using these fuzzing heuristic rules, Orion generates corner-case test inputs for API-level fuzzing. In addition, we extend the seed collection of existing studies to include test inputs for developer APIs. Our evaluation shows that Orion reports 135 vulnerabilities in the latest releases of TensorFlow and PyTorch, 76 of which were confirmed by the library developers. Among the 76 confirmed vulnerabilities, 69 were previously unknown, and 7 have already been fixed. The rest are awaiting further confirmation. For end-user APIs, Orion detected 45.58% and 90% more vulnerabilities in TensorFlow and PyTorch, respectively, compared to the state-of-the-art conventional fuzzer, DeepRel. When compared to the state-of-the-art LLM-based DL fuzzer, AtlasFuz, and Orion detected 13.63% more vulnerabilities in TensorFlow and 18.42% more vulnerabilities in PyTorch. Regarding developer APIs, Orion stands out by detecting 117% more vulnerabilities in TensorFlow and 100% more vulnerabilities in PyTorch compared to the most relevant fuzzer designed for developer APIs, such as FreeFuzz.

Scuzer : A Scheduling Optimization Fuzzer for TVM

The concept of deep learning (DL) compiler was proposed to deploy DL models more efficiently on diverse hardware through optimization techniques. As one of the most popular DL compilers, TVM incorporates three levels (high-level, schedule, and low-level) of optimizations, which can inadvertently introduce code logic bugs and build failure bugs. Among these optimizations, scheduling optimization is the core component of DL compilers, which ensures the acceleration of models on all devices. However, the existing works only focus on the testing of high-level and low-level optimizations in TVM, fail to take the most important and challenging intermediate scheduling optimization layer into consideration. To fill the gap, we propose a Sc heduling optimization oriented f u z zer for TVM, named Scuzer , which is specially designed to effectively detect bugs introduced by the scheduling optimization. In particular, Scuzer first proposes a set of schedule-triggering mutators to actively trigger many scheduling optimizations. Meanwhile, observing that scheduling optimization is closely coupled with program data flow and operator type, Scuzer additionally proposes a set of structure-enriching mutators to enrich the structure of data flows and operators. Based on these carefully designed mutators, Scuzer then devises a multi-objective algorithm that can adaptively select different combinations of objectives at each period to guide the selection of seeds and mutators during fuzzing. We conduct extensive experiments comparing with three state-of-the-art fuzzers that can be applied in testing scheduling optimization to evaluate the effectiveness of Scuzer . The experimental results demonstrate that Scuzer outperforms the 2nd-best state-of-the-art fuzzer by 7.4% in edge coverage and achieves 7 \(\times\) improvement in rule-operator coverage. Scuzer has successfully detected 17 previously unknown bugs (9 are inconsistent results and 5 are inconsistent compilations) in TVM, out of which 10 have been confirmed and 5 been fixed.

Evaluating DL Model Scaling Trade-Offs During Inference via an Empirical Benchmark Analysis

With generative Artificial Intelligence (AI) capturing public attention, the appetite of the technology sector for larger and more complex Deep Learning (DL) models is continuously growing. Traditionally, the focus in DL model development has been on scaling the neural network’s foundational structure to increase computational complexity and enhance the representational expressiveness of the model. However, with recent advancements in edge computing and 5G networks, DL models are now aggressively being deployed and utilized across the cloud–edge–IoT continuum for the realization of in situ intelligent IoT services. This paradigm shift introduces a growing need for AI practitioners, as a focus on inference costs, including latency, computational overhead, and energy efficiency, is long overdue. This work presents a benchmarking framework designed to assess DL model scaling across three key performance axes during model inference: classification accuracy, computational overhead, and latency. The framework’s utility is demonstrated through an empirical study involving various model structures and variants, as well as publicly available datasets for three popular DL use cases covering natural language understanding, object detection, and regression analysis.

Robust thoracic CT image registration with environmental adaptability using dynamic Welsch's function and hierarchical structure-awareness strategy.

Robust registration of thoracic computed tomography (CT) images is strongly impacted by motion during acquisition, high-density objects, and noise, particularly in lower-dose acquisitions. Despite the enhanced registration speed achieved by popular deep learning (DL) methods, their robustness is often neglected. This study aimed to develop a robust thoracic CT image registration algorithm to address the aforementioned issues. A novel, anatomical structure-aware hierarchical registration. By this method, employing a divide-and-conquer approach, dissimilarity metrics, and regularization terms are selected for different regions based on their distinct image features and motion patterns. These terms are then innovatively reconstructed using the Welsch's function, which allows control over the penalty distribution on the loss values. Subsequently, a novel Welsch parameter update strategy is designed for the task of thoracic CT image registration, enabling dynamic sparsity in registration from coarse to fine levels to accommodate various levels of noise and sliding motion. Moreover, the majorization-minimization (MM) algorithm is used to handle the Welsch terms by constructing surrogate functions based on the current variable values for variable update, thereby reducing the complexity of optimization. Experimental results on publicly available deformable image registration lab four-dimensional CT (DIR-Lab 4DCT) and chronic obstructive pulmonary disease (COPD) datasets with and without noise, showed that our proposed method achieves comparable performance to state-of-the-art methods in noise-free scenarios [1.14 and 1.19 mm compared to 1.14 and 1.35 mm target registration errors (TREs)], while demonstrating superior robustness in the presence of noise (1.78 and 2.38 mm compared to 2.00 and 3.31 mm TREs). Ablation studies also validated the effectiveness of each component in the method. A novel and robust algorithm for thoracic CT image registration has been proposed, which has significant potential for valuable clinical applications, including surgical quantitative imaging.

Advancing Bioactivity Prediction through Molecular Docking and Self-Attention.

Bioactivity refers to the ability of a substance to induce biological effects within living systems, often describing the influence of molecules, drugs, or chemicals on organisms. In drug discovery, predicting bioactivity streamlines early-stage candidate screening by swiftly identifying potential active molecules. The popular deep learning methods in bioactivity prediction primarily model the ligand structure-bioactivity relationship under the premise of Quantitative Structure-Activity Relationship (QSAR). However, bioactivity is determined by multiple factors, including not only the ligand structure but also drug-target interactions, signaling pathways, reaction environments, pharmacokinetic properties, and species differences. Our study first integrates drug-target interactions into bioactivity prediction using protein-ligand complex data from molecular docking. We devise a Drug-Target Interaction Graph Neural Network (DTIGN), infusing interatomic forces into intermolecular graphs. DTIGN employs multi-head self-attention to identify native-like binding pockets and poses within molecular docking results. To validate the fidelity of the self-attention mechanism, we gather ground truth data from crystal structure databases. Subsequently, we employ these limited native structures to refine bioactivity prediction via semi-supervised learning. For this study, we establish a unique benchmark dataset for evaluating bioactivity prediction models in the context of protein-ligand complexes, showcasing the superior performance of our method (with an average improvement of 27.03%) through comparison with 9 leading deep learning-based bioactivity prediction methods.

Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models.

In deep learning, different kinds of deep networks typically need different optimizers, which have to be chosen after multiple trials, making the training process inefficient. To relieve this issue and consistently improve the model training speed across deep networks, we propose the ADAptive Nesterov momentum algorithm, Adan for short. Adan first reformulates the vanilla Nesterov acceleration to develop a new Nesterov momentum estimation (NME) method, which avoids the extra overhead of computing gradient at the extrapolation point. Then Adan adopts NME to estimate the gradient's first- and second-order moments in adaptive gradient algorithms for convergence acceleration. Besides, we prove that Adan finds an ϵ-approximate first-order stationary point within O(ϵ-3.5) stochastic gradient complexity on the non-convex stochastic problems (e.g., deep learning problems), matching the best-known lower bound. Extensive experimental results show that Adan consistently surpasses the corresponding SoTA optimizers on vision, language, and RL tasks and sets new SoTAs for many popular networks and frameworks, e.g., ResNet, ConvNext, ViT, Swin, MAE, DETR, GPT-2, Transformer-XL, and BERT. More surprisingly, Adan can use half of the training cost (epochs) of SoTA optimizers to achieve higher or comparable performance on ViT, GPT-2, MAE, etc, and also shows great tolerance to a large range of minibatch size, e.g., from 1 k to 32 k.

Leveraging patent classification based on deep learning: The case study on smart cities and industrial Internet of Things

With the trends of technology convergence and technology interdisciplinarity, technology-field (TF) resolution and classification of patents have gradually been challenged. Whether for patent applicants or for patent examiners, more precisely labeling the TF for a certain patent is important for technological searches. However, determining the TF of a patent may be difficult and may even involve the strategic behavior of patenting, which can cause noise in patent classification systems (PCSs). In addition, some specific patents could contain more TFs than claimed or be assigned questionable IPC codes; subsequently, in a regular search for technology/patents, information could be missed. Considering the advantages of deep learning compared with traditional machine learning algorithms in areas such as natural language processing (NLP), text classification and text sentiment analysis, this paper investigates several popular deep learning models and proposes a large-scale multilabel regression (MLR) model to handle specific patent analyses under situations of small sample learning. To verify the proposed MLR model for patent classification, the case study on smart cities and industrial Internet of Things (IIoT) is conducted. The MLR experiments on the TF resolution of smart cities and IIoT have yielded moderate results compared with those of the latest patent classification studies, which also rely on deep learning and the large language models (LLMs), which include RCNN, Bi-LSTM, BERT and GPT-4 etc. Therefore, the proposed MLR model with a customized loss function could be moderately effective for patent classification within a specific technology theme, could have implications for patent classification and the TF resolution of patents, and could further enrich methodologies for patent mining and informetrics based on artificial intelligence (AI).

AI-Based Noninvasive Blood Glucose Monitoring: Scoping Review.

Current blood glucose monitoring (BGM) methods are often invasive and require repetitive pricking of a finger to obtain blood samples, predisposing individuals to pain, discomfort, and infection. Noninvasive blood glucose monitoring (NIBGM) is ideal for minimizing discomfort, reducing the risk of infection, and increasing convenience. This review aimed to map the use cases of artificial intelligence (AI) in NIBGM. A systematic scoping review was conducted according to the Arksey O'Malley five-step framework. Eight electronic databases (CINAHL, Embase, PubMed, Web of Science, Scopus, The Cochrane-Central Library, ACM Digital Library, and IEEE Xplore) were searched from inception until February 8, 2023. Study selection was conducted by 2 independent reviewers, descriptive analysis was conducted, and findings were presented narratively. Study characteristics (author, country, type of publication, study design, population characteristics, mean age, types of noninvasive techniques used, and application, as well as characteristics of the BGM systems) were extracted independently and cross-checked by 2 investigators. Methodological quality appraisal was conducted using the Checklist for assessment of medical AI. A total of 33 papers were included, representing studies from Asia, the United States, Europe, the Middle East, and Africa published between 2005 and 2023. Most studies used optical techniques (n=19, 58%) to estimate blood glucose levels (n=27, 82%). Others used electrochemical sensors (n=4), imaging (n=2), mixed techniques (n=2), and tissue impedance (n=1). Accuracy ranged from 35.56% to 94.23% and Clarke error grid (A+B) ranged from 86.91% to 100%. The most popular machine learning algorithm used was random forest (n=10) and the most popular deep learning model was the artificial neural network (n=6). The mean overall checklist for assessment of medical AI score on the included papers was 33.5 (SD 3.09), suggesting an average of medium quality. The studies reviewed demonstrate that some AI techniques can accurately predict glucose levels from noninvasive sources while enhancing comfort and ease of use for patients. However, the overall range of accuracy was wide due to the heterogeneity of models and input data. Efforts are needed to standardize and regulate the use of AI technologies in BGM, as well as develop consensus guidelines and protocols to ensure the quality and safety of AI-assisted monitoring systems. The use of AI for NIBGM is a promising area of research that has the potential to revolutionize diabetes management.

Improving the explainability of CNN-LSTM-based flood prediction with integrating SHAP technique

Convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) are popular deep learning architectures currently used for rapid flood simulations. However, deep learning algorithms are difficult to explain, like a “black box” that lacks insight. In order to reveal the intrinsic mechanism of prediction by such architectures, we adopted a coupled CNN-LSTM model based on the explainability technique SHapley Additive exPlanations (SHAP) to predict the rainfall-runoff process and identify key input feature factors, and took the Beijiang River Basin in China as an example, so as to improve the explainability and credibility of this black-box model. The results show that the coupled CNN-LSTM model performs better than the flood predictions compared to the individual CNN or LSTM models under the longest foresight period of 25 h. In particular, the Nash-Sutcliffe Efficiency (NSE) of the former model reaches 0.838, while those of the latter two models are 0.737 and 0.745, respectively. The coupled CNN-LSTM model has a high-accuracy prediction capability, consistently exhibiting NSEs greater than 0.8 for different input time steps and foresight periods. The prediction accuracy is mainly influenced by the observed runoff at the downstream hydrological station from previous time points, while the effects of the input time step length and the foresight period are comparatively negligible. This study provides a new perspective for understanding the potential physical mechanism of black-box models for rainfall-runoff prediction and emphasizes the importance and prospect of the application of explainability techniques.

Exploring the effects of RNNs and deep learning frameworks on real‐time, lightweight, adaptive time series anomaly detection

SummaryReal‐time, lightweight, adaptive time series anomaly detection is increasingly critical in cybersecurity, industrial control, finance, healthcare, and many other domains due to its capability to promptly process time series and detect anomalies without requiring extensive computation resources. While numerous anomaly detection approaches have emerged recently, they generally employ a single type of recurrent neural network (RNN) and are implemented using a single type of deep learning framework. The impacts of using various RNN types across different deep learning frameworks on the performance of these approaches remain unclear due to a lack of comprehensive evaluations. In this article, we aim to investigate the impact of different RNN variants and deep learning frameworks on real‐time, lightweight, and adaptive time series anomaly detection. We reviewed several state‐of‐the‐art anomaly detection approaches and implemented a representative approach using several RNN variants supported by three popular deep learning frameworks. A thorough evaluation was conducted to analyze the detection accuracy, time efficiency, and resource consumption of each implementation using four real‐world, open‐source time series datasets. The results show that RNN variants and deep learning frameworks have a significant impact. Therefore, it is crucial to carefully select appropriate RNN variants and deep learning frameworks for the implementation.

Voxel-wise segmentation for porosity investigation of additive manufactured parts with 3D unsupervised and (deeply) supervised neural networks

Additive Manufacturing (AM) has emerged as a manufacturing process that allows the direct production of samples from digital models. To ensure that quality standards are met in all samples of a batch, X-ray computed tomography (X-CT) is often used in combination with automated anomaly detection. For the latter, deep learning (DL) anomaly detection techniques are increasingly used, as they can be trained to be robust to the material being analysed and resilient to poor image quality. Unfortunately, most recent and popular DL models have been developed for 2D image processing, thereby disregarding valuable volumetric information. Additionally, there is a notable absence of comparisons between supervised and unsupervised models for voxel-wise pore segmentation tasks. This study revisits recent supervised (UNet, UNet++, UNet 3+, MSS-UNet, ACC-UNet) and unsupervised (VAE, ceVAE, gmVAE, vqVAE, RV-VAE) DL models for porosity analysis of AM samples from X-CT images and extends them to accept 3D input data with a 3D-patch approach for lower computational requirements, improved efficiency and generalisability. The supervised models were trained using the Focal Tversky loss to address class imbalance that arises from the low porosity in the training datasets. The output of the unsupervised models was post-processed to reduce misclassifications caused by their inability to adequately represent the object surface. The findings were cross-validated in a 5-fold fashion and include: a performance benchmark of the DL models, an evaluation of the post-processing algorithm, an evaluation of the effect of training supervised models with the output of unsupervised models. In a final performance benchmark on a test set with poor image quality, the best performing supervised model was UNet++ with an average precision of 0.751 ± 0.030, while the best unsupervised model was the post-processed ceVAE with 0.830 ± 0.003. Notably, the ceVAE model, with its post-processing technique, exhibited superior capabilities, endorsing unsupervised learning as the preferred approach for the voxel-wise pore segmentation task.

Detecting marine heatwaves below the sea surface globally using dynamics-guided statistical learning

Extreme warm water events, known as marine heatwaves, cause a variety of adverse impacts on the marine ecosystem. They are occurring more and more frequently across the global ocean. Yet monitoring marine heatwaves below the sea surface is still challenging due to the sparsity of in situ temperature observations. Here, we propose a statistical learning method guided by ocean dynamics and optimal prediction theory, to detect subsurface marine heatwaves based on the observable sea surface temperature and sea surface height. This dynamics-guided statistical learning method shows good skills in detecting subsurface marine heatwaves in the oceanic epipelagic zone over many parts of the global ocean. It outperforms both the classical ordinary least square regression and popular deep learning methods that do not effectively exploit ocean dynamics, with clear dynamical interpretation for its outperformance. Our study provides a useful statistical learning method for near real-time monitoring of subsurface marine heatwaves at a global scale and highlights the importance of exploiting ocean dynamics for enhancing the efficiency and interpretability of statistical learning.

Use of Deep-Learning-Accelerated Gradient Approximation for Reservoir Geological Parameter Estimation

The estimation of space-varying geological parameters is often not computationally affordable for high-dimensional subsurface reservoir modeling systems. The adjoint method is generally regarded as an efficient approach for obtaining analytical gradient and, thus, proceeding with the gradient-based iteration algorithm; however, the infeasible memory requirement and computational demands strictly prohibit its generic implementation, especially for high-dimensional problems. The autoregressive neural network (aNN) model, as a nonlinear surrogate approximation, has gradually received increasing popularity due to significant reduction of computational cost, but one prominent limitation is that the generic application of aNN to large-scale reservoir models inevitably poses challenges in the training procedure, which remains unresolved. To address this issue, model-order reduction could be a promising strategy, which enables us to train the neural network in a very efficient manner. A very popular projection-based linear reduction method, i.e., propel orthogonal decomposition (POD), is adopted to achieve dimensionality reduction. This paper presents an architecture of a projection-based autoregressive neural network that efficiently derives an easy-to-use adjoint model by the use of an auto-differentiation module inside the popular deep learning frameworks. This hybrid neural network proxy, referred to as POD-aNN, is capable of speeding up derivation of reduced-order adjoint models. The performance of POD-aNN is validated through a synthetic 2D subsurface transport model. The use of POD-aNN significantly reduces the computation cost while the accuracy remains. In addition, our proposed POD-aNN can easily obtain multiple posterior realizations for uncertainty evaluation. The developed POD-aNN emulator is a data-driven approach for reduced-order modeling of nonlinear dynamic systems and, thus, should be a very efficient modeling tool to address many engineering applications related to intensive simulation-based optimization.

WhiteFox: White-Box Compiler Fuzzing Empowered by Large Language Models

Compiler correctness is crucial, as miscompilation can falsify program behaviors, leading to serious consequences over the software supply chain. In the literature, fuzzing has been extensively studied to uncover compiler defects. However, compiler fuzzing remains challenging: Existing arts focus on black- and grey-box fuzzing, which generates test programs without sufficient understanding of internal compiler behaviors. As such, they often fail to construct test programs to exercise intricate optimizations. Meanwhile, traditional white-box techniques, such as symbolic execution, are computationally inapplicable to the giant codebase of compiler systems. Recent advances demonstrate that Large Language Models (LLMs) excel in code generation/understanding tasks and even have achieved state-of-the-art performance in black-box fuzzing. Nonetheless, guiding LLMs with compiler source-code information remains a missing piece of research in compiler testing. To this end, we propose WhiteFox, the first white-box compiler fuzzer using LLMs with source-code information to test compiler optimization, with a spotlight on detecting deep logic bugs in the emerging deep learning (DL) compilers. WhiteFox adopts a multi-agent framework: (i) an LLM-based analysis agent examines the low-level optimization source code and produces requirements on the high-level test programs that can trigger the optimization; (ii) an LLM-based generation agent produces test programs based on the summarized requirements. Additionally, optimization-triggering tests are also used as feedback to further enhance the test generation prompt on the fly. Our evaluation on the three most popular DL compilers (i.e., PyTorch Inductor, TensorFlow-XLA, and TensorFlow Lite) shows that WhiteFox can generate high-quality test programs to exercise deep optimizations requiring intricate conditions, practicing up to 8 times more optimizations than state-of-the-art fuzzers. To date, WhiteFox has found in total 101 bugs for the compilers under test, with 92 confirmed as previously unknown and 70 already fixed. Notably, WhiteFox has been recently acknowledged by the PyTorch team, and is in the process of being incorporated into its development workflow. Finally, beyond DL compilers, WhiteFox can also be adapted for compilers in different domains, such as LLVM, where WhiteFox has already found multiple bugs.

A systematic review of deep learning techniques for plant diseases

Agriculture is one of the most crucial sectors, meeting the fundamental food needs of humanity. Plant diseases increase food economic and food security concerns for countries and disrupt their agricultural planning. Traditional methods for detecting plant diseases require a lot of labor and time. Consequently, many researchers and institutions strive to address these issues using advanced technological methods. Deep learning-based plant disease detection offers considerable progress and hope compared to classical methods. When trained with large and high-quality datasets, these technologies robustly detect diseases on plant leaves in early stages. This study systematically reviews the application of deep learning techniques in plant disease detection by analyzing 160 research articles from 2020 to 2024. The studies are examined in three different areas: classification, detection, and segmentation of diseases on plant leaves, while also thoroughly reviewing publicly available datasets. This systematic review offers a comprehensive assessment of the current literature, detailing the most popular deep learning architectures, the most frequently studied plant diseases, datasets, encountered challenges, and various perspectives. It provides new insights for researchers working in the agricultural sector. Moreover, it addresses the major challenges in the field of disease detection in agriculture. Thus, this study offers valuable information and a suitable solution based on deep learning applications for agricultural sustainability.

Classification Model of Grassland Desertification Based on Deep Learning

Grasslands are one of the most important ecosystems on earth, and the impact of grassland desertification on the earth’s environment and ecosystem cannot be ignored. Accurately distinguishing grassland desertification types has important application value. The appropriate grazing strategies can be implemented based on these distinctions. Grassland conservation measures can be tailored accordingly. This contributes to further protecting and restoring grassland vegetation. This project takes color images labeled with the desertification types of grasslands as the research object, uses the currently popular deep learning model as the classification tool, and then establishes a color image-based grassland desertification classification model based on the feature extraction network, based on the Vision Transformer model, by comparing the various deep learning image classification models. The experimental results show that, despite the complex structure and large number of parameters of the grassland desertification classification model obtained in this project, the test accuracy rate reaches 88.72% and the training loss is only 0.0319. Compared with the popular classification models such as VGG16, ResNet50, ResNet101, DenseNet101, DenseNet169, and DenseNet201, and so on, the Vision Transformer demonstrates clear advantages in classification accuracy, fitting ability, and generalization capacity. By integrating with deep learning technology, the model can be applied to grassland management and ecological restoration. Mobile devices can be used to conveniently capture image data, and information can be processed quickly. This provides efficient tools for grazing managers, environmental scientists, and conservation organizations. These tools assist in quickly assessing the extent of grassland desertification, optimizing grassland management and conservation decisions. Furthermore, strong technical support is offered for the ecological restoration and sustainable management of desertification grasslands.

A systematic review on deep learning based methods for cervical cell image analysis

Cervical cytology image analysis is indispensable for the detection of abnormal cervical cells. Traditionally, manual screening is time-consuming and labor-intensive. Therefore, a lot of deep learning (DL)-based automatic detection methods have been employed in this field to provide timely, accurate and objective results. In this study, we systematically review the current developments in cervical cell image analysis with DL methods. Specifically, we first present the most popular DL models that are widely applied in cervical cell analysis. Second, we describe the methodology for conducting this review. Third, we provide all publicly available datasets related to cervical cell images to the best of our knowledge. Then, we introduce relevant evaluation metrics and loss functions. Next, we summarize and assort the applications for cervical cell classification and segmentation. Afterwards, we discuss about current challenges and future research directions in this field. Finally, we draw the conclusion of this review.According to the analysis, we conclude that the studies based on DL models have maintained an increasing trend in recent years, which indicates the potential of DL in cervical cell image analysis. In cervical cell image classification, CNN is the most commonly used DL model. Among CNN models, we can find that VGGNet and ResNet are the most popular network architectures for the classification of cervical cells. Transformer is the second commonly used DL model. Moreover, Herlev and SIPaKMeD are the most popular public datasets used for cervical cell classification. In cervical cell segmentation, U-Net and FCN are the two most popular DL architectures. In addition, ISBI2014 and Herlev datasets are the most frequently used among the existing publicly available segmentation datasets. However, there are some issues in this field, such as poor cervical cell classification performance as a result of similar pathological properties between different cell categories. Therefore, it is necessary to develop more effective methods with DL models to improve these issues in the future research.

A Comparative Study of Deep Learning in Breast Ultrasound Lesion Detection: From Two-Stage to One-Stage, from Anchor-Based to Anchor-Free

Article A Comparative Study of Deep Learning in Breast Ultrasound Lesion Detection: From Two-Stage to One-Stage, from Anchor-Based to Anchor-Free Yu Wang 1, Qi Zhao 1, Baihua Zhang 2, Dingcheng Tian 1, Ruyi Zhang 1 and Wan Zhong 3,∗ 1 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang 110024 , China 2 Research Institute for Medical and Biological Engineering, Ningbo University, Ningbo 315000, China 3 General Hospital of Northern Theater Command, Shenyang 110024, China ∗ Correspondence: wzhong_88@163.com Received: 16 July 2024; Revised: 26 August 2024; Accepted: 27 August 2024; Published: 4 September 2024 Abstract: Breast cancer is one of the most common tumors among women in the world, and its early screening is crucial to improve the survival rate of patients. Breast ultrasound, with the characteristics of non radiation, real-time imaging and easy operation, has become a common method for breast cancer detection. However, this method has some problems, such as low imaging quality and strong subjectivity of diagnosis results, which affect the accurate diagnosis of breast cancer. With the ongoing advancement of deep learning technology, intelligent breast cancer detection systems have effectively overcome these challenges, enhancing diagnostic accuracy and efficiency. This study uses nine popular deep learning object detection networks (including two-stage, one-stage, anchor-based, and anchor-free networks) for the detection of breast lesions and compares the results of these methods. The experiments show that the anchor-based Single Shot MultiBox Detector (SSD) network excels in overall performance, while the anchor-free Fully Convolutional One-stage Object Detector (FCOS) exhibits the best generalization ability. Moreover, the results also indicate that, in the context of breast lesion detection, anchor-based networks generally outperform anchor-free networks.