High Generalization Capability Research Articles

3DVascNet: An Automated Software for Segmentation and Quantification of Mouse Vascular Networks in 3D.

Analysis of vascular networks is an essential step to unravel the mechanisms regulating the physiological and pathological organization of blood vessels. So far, most of the analyses are performed using 2-dimensional projections of 3-dimensional (3D) networks, a strategy that has several obvious shortcomings. For instance, it does not capture the true geometry of the vasculature and generates artifacts on vessel connectivity. These limitations are accepted in the field because manual analysis of 3D vascular networks is a laborious and complex process that is often prohibitive for large volumes. To overcome these issues, we developed 3DVascNet, a deep learning-based software for automated segmentation and quantification of 3D retinal vascular networks. 3DVascNet performs segmentation based on a deep learning model, and it quantifies vascular morphometric parameters such as vessel density, branch length, vessel radius, and branching point density. We tested the performance of 3DVascNet using a large data set of 3D microscopy images of mouse retinal blood vessels. We demonstrated that 3DVascNet efficiently segments vascular networks in 3D and that vascular morphometric parameters capture phenotypes detected by using manual segmentation and quantification in 2 dimension. In addition, we showed that, despite being trained on retinal images, 3DVascNet has high generalization capability and successfully segments images originating from other data sets and organs. Overall, we present 3DVascNet, a freely available software that includes a user-friendly graphical interface for researchers with no programming experience, which will greatly facilitate the ability to study vascular networks in 3D in health and disease. Moreover, the source code of 3DVascNet is publicly available, thus it can be easily extended for the analysis of other 3D vascular networks by other users.

Dual-branch information extraction and local attention anchor-free network for defect detection

In the production process, the presence of surface defects seriously affects the quality of industrial products. Existing defect detectors are not suitable for surface with scattered distribution and complex texture of defects. In this study, a dual-branch information extraction and local attention anchor-free network for defect detection (DLA-FCOS), which is based on the fully convolutional one-stage network, is proposed to accurately locate and detect surface defects of industrial products. Firstly, a dual-branch feature extraction network (DFENeT) is proposed and used to improve the extraction ability of complex defects. Then, a local feature enhancement module is proposed, and a residual connection is established to enrich local semantic information. Meanwhile, the self-attention mechanism is introduced to form local attentional residual feature pyramid networks (LA-RFPN) to eliminate the influences of feature misalignments. The mean average accuracy (mAP) and frames per second (FPS) of the proposed DLA-FCOS on the cut layer of the tobacco packet defect dataset (CLTP-DD) are 96.8% and 20.7, respectively, which meets the requirements for accurate and real-time defect detection. Meanwhile, the average accuracy of the proposed DLA-FCOS on the NEU-DET and GC10-DET datasets is 78.4% and 67.7%, respectively. The results demonstrate that the DLA-FCOS has good feasibility and high generalization capability to perform defect detection tasks of industrial products.

Training and Comparison of nnU-Net and DeepMedic Methods for Autosegmentation of Pediatric Brain Tumors.

Tumor segmentation is essential in surgical and treatment planning and response assessment and monitoring in pediatric brain tumors, the leading cause of cancer-related death among children. However, manual segmentation is time-consuming and has high interoperator variability, underscoring the need for more efficient methods. After training, we compared 2 deep-learning-based 3D segmentation models, DeepMedic and nnU-Net, with pediatric-specific multi-institutional brain tumor data based on multiparametric MR images. Multiparametric preoperative MR imaging scans of 339 pediatric patients (n = 293 internal and n = 46 external cohorts) with a variety of tumor subtypes were preprocessed and manually segmented into 4 tumor subregions, ie, enhancing tumor, nonenhancing tumor, cystic components, and peritumoral edema. After training, performances of the 2 models on internal and external test sets were evaluated with reference to ground truth manual segmentations. Additionally, concordance was assessed by comparing the volume of the subregions as a percentage of the whole tumor between model predictions and ground truth segmentations using the Pearson or Spearman correlation coefficients and the Bland-Altman method. The mean Dice score for nnU-Net internal test set was 0.9 (SD, 0.07) (median, 0.94) for whole tumor; 0.77 (SD, 0.29) for enhancing tumor; 0.66 (SD, 0.32) for nonenhancing tumor; 0.71 (SD, 0.33) for cystic components, and 0.71 (SD, 0.40) for peritumoral edema, respectively. For DeepMedic, the mean Dice scores were 0.82 (SD, 0.16) for whole tumor; 0.66 (SD, 0.32) for enhancing tumor; 0.48 (SD, 0.27) for nonenhancing tumor; 0.48 (SD, 0.36) for cystic components, and 0.19 (SD, 0.33) for peritumoral edema, respectively. Dice scores were significantly higher for nnU-Net (P ≤ .01). Correlation coefficients for tumor subregion percentage volumes were higher (0.98 versus 0.91 for enhancing tumor, 0.97 versus 0.75 for nonenhancing tumor, 0.98 versus 0.80 for cystic components, 0.95 versus 0.33 for peritumoral edema in the internal test set). Bland-Altman plots were better for nnU-Net compared with DeepMedic. External validation of the trained nnU-Net model on the multi-institutional Brain Tumor Segmentation Challenge in Pediatrics (BraTS-PEDs) 2023 data set revealed high generalization capability in the segmentation of whole tumor, tumor core (a combination of enhancing tumor, nonenhancing tumor, and cystic components), and enhancing tumor with mean Dice scores of 0.87 (SD, 0.13) (median, 0.91), 0.83 (SD, 0.18) (median, 0.89), and 0.48 (SD, 0.38) (median, 0.58), respectively. The pediatric-specific data-trained nnU-Net model is superior to DeepMedic for whole tumor and subregion segmentation of pediatric brain tumors.

Convolutional Forecasting of Particulate Matter: Toward a Data-Driven Generalized Model

Air pollution poses a significant threat to human health and ecosystems. Forecasting the concentration of key pollutants like particulate matter can help support air quality planning and prevention measures. Deep learning methods are becoming increasingly popular for predicting air pollution and particulate matter concentration. Architectures like Convolutional Neural Networks can effectively account for the geographical features of the study domain. This work tests a Feed-Forward, a Long Short-Term Memory (LSTM), and a Convolutional Neural Network (CNN) on a polluted geographical domain in northern Italy. The best convolutional architecture was then implemented in two other quite different regions. The results show that the same CNN architecture provides remarkably accurate forecasts in all applications and that a network trained on PM10 data can accurately forecast PM2.5 concentrations up to 10 days ahead. These results suggest that the proposed CNN has high generalization capabilities and can thus be reliably used as a forecasting model for different areas.

Prioritizing Causation in Decision Trees: A Framework for Interpretable Modeling

As a popular machine learning model, decision trees classify and generalize well, but face challenges in engineering applications: 1) Sensitivity to perturbations and lack of interpretability due to correlation reliance. 2) Manual setting of stopping criterion which is unrelated to correlation strength and easily leads to over-partitioning. To address these two challenges, we first theoretically analyze what leads to sub-optimal decision trees. By incorporating causal discovery, this limitation can be attributed to the fact that trees grown with spurious correlations often fall into sub-optimal that lead to overfitting and unfair behaviors. Neglecting causality motivates us to develop a ‘better’ tree with low Kolmogorov complexity and high generalization capability. Then we propose a causality decision tree framework, CausalDT, based on our theoretical expectation, where Hilbert-Schmidt independence criterion serves as a baseline. Unlike previous approaches that prioritize relevance, our framework determines branch nodes based on causation between features, with the significance level determining whether the tree should be expanded further. Experimental results demonstrate that our model maintains performance while reducing average tree depth by 35% on various datasets. Furthermore, our model enhances decision fairness and interpretability.

Prediction of chemical reaction yields with large-scale multi-view pre-training

Developing machine learning models with high generalization capability for predicting chemical reaction yields is of significant interest and importance. The efficacy of such models depends heavily on the representation of chemical reactions, which has commonly been learned from SMILES or graphs of molecules using deep neural networks. However, the progression of chemical reactions is inherently determined by the molecular 3D geometric properties, which have been recently highlighted as crucial features in accurately predicting molecular properties and chemical reactions. Additionally, large-scale pre-training has been shown to be essential in enhancing the generalization capability of complex deep learning models. Based on these considerations, we propose the Reaction Multi-View Pre-training (ReaMVP) framework, which leverages self-supervised learning techniques and a two-stage pre-training strategy to predict chemical reaction yields. By incorporating multi-view learning with 3D geometric information, ReaMVP achieves state-of-the-art performance on two benchmark datasets. Notably, the experimental results indicate that ReaMVP has a significant advantage in predicting out-of-sample data, suggesting an enhanced generalization ability to predict new reactions. Scientific Contribution: This study presents the ReaMVP framework, which improves the generalization capability of machine learning models for predicting chemical reaction yields. By integrating sequential and geometric views and leveraging self-supervised learning techniques with a two-stage pre-training strategy, ReaMVP achieves state-of-the-art performance on benchmark datasets. The framework demonstrates superior predictive ability for out-of-sample data and enhances the prediction of new reactions.

Hybrid load prediction model of 5G base station based on time series decomposition and GRU network with parameter optimization

AbstractTo ensure the safe and stable operation of 5G base stations, it is essential to accurately predict their power load. However, current short‐term prediction methods are rarely applied rationally in pertinent circumstances to the features of base station power load over time. For high accuracy and generalization capabilities, this work proposes a hybrid approach that combines gated recurrent unit (GRU) with particle swarm optimization (PSO) and completes ensemble empirical mode decomposition with adaptive noise (CEEMDAN). The complex load is decomposed into multiple components to improve the fit of the neural network. To deal with the nonlinearity that constrains the modal aliasing and data noise, CEEMDAN is used to reconstruct the trend and noise sequences. The GRU network is utilized to improve the neural network fitness and obtain long‐term features. In addition, PSO is used to optimize the number of neurons and the learning rate of the GRU. Compared with existing neural network GRUs, the RMSE is reduced by 16.6, the mean absolute error is reduced by 11.97, and the coefficient of R2 is improved by 0.13, indicating that the model has a better fitting effect. The comparisons prove that the proposed model has better accuracy than the existing methods.

Research on the prediction of mine water inrush disasters based on multi-factor spatial game reconstruction

Water damage in mines poses a widespread challenge in the coal mining industry. Gaining a comprehensive understanding of the multi-factor spatial catastrophe evolution mechanism and process of floor water inrush is crucial, which will enable the achievement of dynamic, quantitative, and precise early warning systems. It holds significant theoretical guidance for implementing effective water prevention and control measures in coal mines. This study focuses on the issue of water inrush in the coal seam floor, specifically in the context of Pengzhuang coal mine. By utilizing a small sample of non-linear characteristics derived from drilling geological data, we adopt a multifactor spatial perspective that considers geological structure and hydrogeological conditions. In light of this, we propose a quantitative risk prediction model that integrates the coupled theoretical analysis, statistical analysis, and machine learning simulation methods. Firstly, the utilization of a quantification approach employing a triangular fuzzy number allows for the representation of a comparative matrix based on empirical values. Simultaneously, the networked risk transmission effect of underlying control risk factors is taken into consideration. The application of principal component analysis optimizes the entropy weight method, effectively reducing the interference caused by multifactor correlation. By employing game theory, the subjective and objective weight proportions of the control factors are reasonably allocated, thereby establishing a vulnerability index model based on a comprehensive weighting of subjective and objective factors. Secondly, the WOA-RF-GIS approach is employed to comprehensively explore the interconnectedness of water diversion channel data. Collaborative Kriging interpolation is utilized to enhance the dimensionality of the data and facilitate spatial information processing. Lastly, the representation of risk is coupled with necessary and sufficient condition layers, enabling the qualitative visualization of quantitative results. This approach aims to accurately predict disaster risk with limited sample data, ultimately achieving the goal of precise risk assessment. The research findings demonstrate that the reconstructed optimization model based on multi-factor spatial game theory exhibits high precision and generalization capability. This model effectively unveils the non-linear dynamic processes associated with floor water inrush, which are influenced by multiple factors, characterized by limited data volume, and governed by complex formation mechanisms. The identification of high-risk areas for water inrush is achieved with remarkable accuracy, providing invaluable technical support for the formulation of targeted water prevention and control measures, ultimately ensuring the safety of coal mining operations.

Emulator of PR‐DNS: Accelerating Dynamical Fields With Neural Operators in Particle‐Resolved Direct Numerical Simulation

AbstractParticle‐resolved direct numerical simulations (PR‐DNS) play an increasing role in investigating aerosol‐cloud‐turbulence interactions at the most fundamental level of processes. However, the high computational cost associated with high resolution simulations poses considerable challenges for large domain or long duration simulation using PR‐DNS. To address these issues, here we present an emulator of the complex physics‐based PR‐DNS developed by use of the data‐driven Fourier Neural Operator (FNO) method. The effectiveness of the method is showcased by presenting turbulence and temperature fields in a two‐dimensional space. The results demonstrate high accuracy at various resolutions and the emulator is two orders of magnitude cheaper in terms of computational demand compared to the physics‐based PR‐DNS model. Furthermore, the FNO emulator exhibits strong generalization capabilities for different initial conditions and ultra‐high‐resolution without the need to retrain models. These findings highlight the potential of the FNO method as a promising tool to simulate complex fluid dynamics problems with high accuracy, computational efficiency, and generalization capabilities, enhancing our understanding of the aerosol‐cloud‐precipitation system.

The Improved U-STFM: A Deep Learning-Based Nonlinear Spatial-Temporal Fusion Model for Land Surface Temperature Downscaling

The thermal band of a satellite platform enables the measurement of land surface temperature (LST), which captures the spatial-temporal distribution of energy exchange between the Earth and the atmosphere. LST plays a critical role in simulation models, enhancing our understanding of physical and biochemical processes in nature. However, the limitations in swath width and orbit altitude prevent a single sensor from providing LST data with both high spatial and high temporal resolution. To tackle this challenge, the unmixing-based spatiotemporal fusion model (STFM) offers a promising solution by integrating data from multiple sensors. In these models, the surface reflectance is decomposed from coarse pixels to fine pixels using the linear unmixing function combined with fractional coverage. However, when downsizing LST through STFM, the linear mixing hypothesis fails to adequately represent the nonlinear energy mixing process of LST. Additionally, the original weighting function is sensitive to noise, leading to unreliable predictions of the final LST due to small errors in the unmixing function. To overcome these issues, we selected the U-STFM as the baseline model and introduced an updated version called the nonlinear U-STFM. This new model incorporates two deep learning components: the Dynamic Net (DyNet) and the Chang Ratio Net (RatioNet). The utilization of these components enables easy training with a small dataset while maintaining a high generalization capability over time. The MODIS Terra daytime LST products were employed to downscale from 1000 m to 30 m, in comparison with the Landsat7 LST products. Our results demonstrate that the new model surpasses STARFM, ESTARFM, and the original U-STFM in terms of prediction accuracy and anti-noise capability. To further enhance other STFMs, these two deep-learning components can replace the linear unmixing and weighting functions with minor modifications. As a deep learning-based model, it can be pretrained and deployed for online prediction.

Research on 3D Geological Modeling Method Based on Deep Neural Networks for Drilling Data

Three-dimensional (3D) models provide the most intuitive representation of geological conditions. Traditional modeling methods heavily depend on technicians’ expertise and lack ease of updating. In this study, we introduce a deep learning-based method for 3D geological implicit modeling, leveraging a substantial dataset of geological drilling data. By applying resampling and normalization techniques, we standardize drilling data and significantly expand the dataset, making it suitable for training deep neural networks. Utilizing the characteristics of the sample data, we design and establish the network structure, loss function, and parameter configurations, resulting in the training of a deep neural network with high accuracy and robust generalization capability. Ultimately, we utilize the dataset generated from the network’s predictions to render and construct the 3D geological model. The research in this paper demonstrates the significant promise of deep neural networks in addressing geological challenges. The deep learning-based implicit 3D modeling method surpasses traditional approaches in terms of generalization, convenience, and adaptability.

An Improved Model Combining Machine Learning and Kalman Filtering Architecture for State of Charge Estimation of Lithium-Ion Batteries

Accurate state of charge (SOC) estimation of lithium-ion batteries is a fundamental prerequisite for ensuring the normal and safe operation of electric vehicles, and it is also a key technology component in battery management systems. In recent years, lithium-ion battery SOC estimation methods based on data-driven approaches have gained significant popularity. However, these methods commonly face the issue of poor model generalization and limited robustness. To address such issues, this study proposes a closed-loop SOC estimation method based on simulated annealing-optimized support vector regression (SA-SVR) combined with minimum error entropy based extended Kalman filter (MEE-EKF) algorithm. Firstly, a probability-based SA algorithm is employed to optimize the internal parameters of the SVR, thereby enhancing the precision of original SOC estimation. Secondly, utilizing the framework of the Kalman filter, the optimized SVR results are incorporated as the measurement equation and further processed through the MEE-EKF, while the ampere-hour integral physical model serves as the state equation, effectively attenuating the measurement noise, enhancing the estimation accuracy, and improving generalization ability. The proposed method is validated through battery testing experiments conducted under three typical operating conditions and one complex and random operating condition with wide temperature variations under only one condition training. The results demonstrate that the proposed method achieves a mean absolute error below 0.60% and a root mean square error below 0.73% across all operating conditions, showcasing a significant improvement in estimation accuracy compared to the benchmark algorithms. The high precision and generalization capability of the proposed method are evident, ensuring accurate SOC estimation for electric vehicles.

Deep Feature Statistics Mapping for Generalized Screen Content Image Quality Assessment.

The statistical regularities of natural images, referred to as natural scene statistics, play an important role in no-reference image quality assessment. However, it has been widely acknowledged that screen content images (SCIs), which are typically computer generated, do not hold such statistics. Here we make the first attempt to learn the statistics of SCIs, based upon which the quality of SCIs can be effectively determined. The underlying mechanism of the proposed approach is based upon the mild assumption that the SCIs, which are not physically acquired, still obey certain statistics that could be understood in a learning fashion. We empirically show that the statistics deviation could be effectively leveraged in quality assessment, and the proposed method is superior when evaluated in different settings. Extensive experimental results demonstrate the Deep Feature Statistics based SCI Quality Assessment (DFSS-IQA) model delivers promising performance compared with existing NR-IQA models and shows a high generalization capability in the cross-dataset settings. The implementation of our method is publicly available at https://github.com/Baoliang93/DFSS-IQA.

Improving domain generalization by hybrid domain attention and localized maximum sensitivity

Domain generalization has attracted much interest in recent years due to its practical application scenarios, in which the model is trained using data from various source domains but is tested using data from an unseen target domain. Existing domain generalization methods concern all visual features, including irrelevant ones with the same priority, which easily results in poor generalization performance of the trained model. In contrast, human beings have strong generalization capabilities to distinguish images from different domains by focusing on important features while suppressing irrelevant features with respect to labels. Motivated by this observation, we propose a channel-wise and spatial-wise hybrid domain attention mechanism to force the model to focus on more important features associated with labels in this work. In addition, models with higher robustness with respect to small perturbations of inputs are expected to have higher generalization capability, which is preferable in domain generalization. Therefore, we propose to reduce the localized maximum sensitivity of the small perturbations of inputs in order to improve the network’s robustness and generalization capability. Extensive experiments on PACS, VLCS, and Office-Home datasets validate the effectiveness of the proposed method.

An adaptive physics-informed deep learning method for pore pressure prediction using seismic data

Accurate prediction of formation pore pressure is essential to predict fluid flow and manage hydrocarbon production in petroleum engineering. Recent deep learning technique has been receiving more interest due to the great potential to deal with pore pressure prediction. However, most of the traditional deep learning models are less efficient to address generalization problems. To fill this technical gap, in this work, we developed a new adaptive physics-informed deep learning model with high generalization capability to predict pore pressure values directly from seismic data. Specifically, the new model, named CGP-NN, consists of a novel parametric features extraction approach (1DCPP), a stacked multilayer gated recurrent model (multilayer GRU), and an adaptive physics-informed loss function. Through machine training, the developed model can automatically select the optimal physical model to constrain the results for each pore pressure prediction. The CGP-NN model has the best generalization when the physics-related metric λ=0.5. A hybrid approach combining Eaton and Bowers methods is also proposed to build machine-learnable labels for solving the problem of few labels. To validate the developed model and methodology, a case study on a complex reservoir in Tarim Basin was further performed to demonstrate the high accuracy on the pore pressure prediction of new wells along with the strong generalization ability. The adaptive physics-informed deep learning approach presented here has potential application in the prediction of pore pressures coupled with multiple genesis mechanisms using seismic data.

A Data-Driven Approach to Predict the ROP of Deep Wells in Fukang Sag

In the deep well drilling process in the Fukang Depression of the Eastern Junggar Basin, rock fracturing issues and low rate of penetration (ROP) have posed significant challenges to drilling efficiency. Accurate predictions of ROP prior to drilling are of considerable value in this context. Precise predictions enable on-site engineering teams to proactively identify drilling difficulties and anticipate potential complex scenarios, facilitating them in designing preventive measures in advance, such as selecting appropriate drill bits, adjusting drilling parameters, or employing specific drilling techniques to address these issues. This, in turn, enhances drilling efficiency and greatly reduces drilling risks. Traditional mechanical-specific energy drilling rate models, despite their widespread use, exhibit significant disparities with actual results when predicting ROP. These models only consider the influence of drill bits, drilling tools, and some drilling parameters on ROP, failing to adequately account for the variations caused by engineering factors and failing to capture the interrelationships between various parameters, especially when dealing with complex subsurface formations in the Fukang Depression. Random forest is a non-parametric algorithm in the field of machine learning that is suitable for analyzing and predicting ROP affected by various complex and non-linear drilling parameters. This paper establishes a Random forest model based on a dataset containing multiple variables of logging parameters and the actual ROP. The model ranks and assesses the important feature parameters of ROP to reveal their impact. Additionally, the model uses bootstrap sampling and feature random selection to construct multiple decision trees, reducing the risk of overfitting and endowing the model with a high generalization capability. Evaluation metrics indicate that the model exhibits a high prediction accuracy and performs well, significantly improving the accuracy of mechanical drilling rate predictions in the deep wells of the Fukang Depression. This model provides robust support and serves as a positive demonstration for addressing mechanical drilling rate issues in complex subsurface formations in the future.

TransPV: Refining photovoltaic panel detection accuracy through a vision transformer-based deep learning model

The increasing need to develop renewable energy sources to combat climate change has led to a significant rise in demand for photovoltaic (PV) installations. Consequently, accurately detecting and estimating the capacity and potential for electricity generation of these installed PV systems has become crucial for effective energy management. However, due to the limitation of localized mechanism in convolution operations, traditional FCN-based methods face challenges in capturing global dependencies among remote sensing image patches. As a result, they struggle to model long-range interactions between different PV panels with diverse structures, including varying size, shape, and texture, particularly in distributed residential areas. To tackle the challenge of modeling PV panels with diverse structures, we propose a coupled U-Net and Vision Transformer model named TransPV for refining PV semantic segmentation. Specifically, Mix Transformer block is incorporated in the encoder to enhance the modeling of global context, while the U-Shaped structure enables the combination of multi-level features, resulting in enriched feature representation, which helps the model better understand objects in the images, regardless of their shape, size, and texture. In addition, we implement the PointRend module in the decoder to obtain finer segmentation boundary details, and utilize a novel refiner loss function during the training process to alleviate the problem of extremely unbalanced samples. The experimental results from the Heilbronn datasets demonstrate the remarkable performance of our proposed TransPV model in addressing intra-class diversity of PV structures, surpassing previous state-of-the-art methods with an impressive IoU and accuracy of 0.802 and 0.876, respectively. Furthermore, our model exhibits high generalization capability on the BDPV dataset with IoU of 0.745. The obtained results highlight the superiority of TransPV in improving accuracy and addressing diverse structure issues in PV segmentation, which provides valuable insights for optimizing the efficiency and sustainability of renewable energy.

A lightweight and scalable greenhouse mapping method based on remote sensing imagery

Seeking a low-cost, high-efficiency greenhouse mapping technology has immense significance. While greenhouse extraction methods using deep learning have been proposed, the challenge of extracting dense small objects remains an unresolved problem. The inherent downscaling strategy in general-purpose semantic segmentation (SS) models renders them unsuitable for such tasks. In contrast, the dramatically increasing computational complexity associated with this problem may result in an unaffordable cost for consumer-level applications. To address the aforementioned challenges, this study presents a novel greenhouse mapping model based on remote sensing (RS) images, which not only exhibits high precision and robust generalization capabilities but also offers significant lightweight advantages. To meet broader needs, we also provide corresponding customizable and scalable rules that allow for a trade-off between accuracy and speed. To evaluate the performance of our model, we select several representative works to conduct benchmark experiments on a self-annotated dataset. The results demonstrate that our method can provide more powerful visual representations for greenhouse segmentation with minimal cost. Compared to the control group, the proposed method achieves an mIoU improvement of 1.116 %-10.77 % using only 3.282 M parameters, while maintaining a considerable inference speed. Code will be available at: https://github.com/W-qp/EGENet.git

Simplifying YOLOv5 for deployment in a real crop monitoring setting

Deep learning-based object detection models have become a preferred choice for crop detection tasks in crop monitoring activities due to their high accuracy and generalization capabilities. However, their high computational demand and large memory footprint pose a challenge for use on mobile embedded devices deployed in crop monitoring settings. Various approaches have been taken to minimize the computational cost and reduce the size of object detection models such as channel and layer pruning, detection head searching, backbone optimization, etc. In this work, we approached computational lightening, model compression, and speed improvement by discarding one or more of the three detection scales of the YOLOv5 object detection model. Thus, we derived up to five separate fast and light models, each with only one or two detection scales. To evaluate the new models for a real crop monitoring use case, the models were deployed on NVIDIA Jetson nano and NVIDIA Jetson Orin devices. The new models achieved up to 21.4% reduction in giga floating-point operations per second (GFLOPS), 31.9% reduction in number of parameters, 30.8% reduction in model size, 28.1% increase in inference speed, with only a small average accuracy drop of 3.6%. These new models are suitable for crop detection tasks since the crops are usually of similar sizes due to the high likelihood of being in the same growth stage, thus, making it sufficient to detect the crops with just one or two detection scales.

Automatic delineation of laryngeal squamous cell carcinoma during endoscopy

White Light (WL) and Narrow Band Imaging (NBI) endoscopy are widely used to assess the superficial spreading of laryngeal squamous cell carcinoma (LSCC). However, the analysis of images requires a high level of attention and extensive clinical expertise, leading to inter-clinician variability on the assessment of tumor margins. Computer-aided segmentation can automate the identification of LSCC margins, supporting clinicians in this challenging task. In this paper, we present SegMENT-Plus, a Deep Learning segmentation convolutional network specifically developed and optimized for accurate delineation of LSCC. SegMENT-Plus uses EfficienstNetB5 as encoder with a new modified Atrous Spatial Pyramid Pooling (m-ASPP) block that integrates Channel Block Attention Module (CBAM) and Squeeze Excitation (SE). In this new architecture, CBAM extracts local and global LSCC features from the encoder, while the SE block refines cancer segmentation on each dilated convolution output. SegMENT-Plus was trained and evaluated on a multi-center dataset including clinical data from three different hospitals. A total of 4289 annotated laryngeal images from 766 patients were included in this study. The experiments showed that SegMENT-Plus achieved a Dice Similarity Coefficient (DSC) between 81.4% and 84.9% and an Intersection over Union (IOU) between 81.8% and 85.7% on the data from the different hospitals, attesting its high performance and generalization capability. The proposed segmentation architecture also demonstrated statistically significant improvement in DSC and IoU compared to other state of the art architectures, showing that this work is a concrete foundation towards a clinical system for the automatic delineation of LSCC margins in endoscopic images.