Convolutional Neural Network Training Research Articles

Predicting Cloud-To-Ground Lightning in the Western United States From the Large-Scale Environment Using Explainable Neural Networks.

Lightning is a major source of wildfire ignition in the western United States (WUS). We build and train convolutional neural networks (CNNs) to predict the occurrence of cloud-to-ground (CG) lightning across the WUS during June-September from the spatial patterns of seven large-scale meteorological variables from reanalysis (1995-2022). Individually trained CNN models at each 1°×1° grid cell (n=285 CNNs) show high skill at predicting CG lightning days across the WUS (median AUC=0.8) and perform best in parts of the interior Southwest where summertime CG lightning is most common. Further, interannual correlation between observed and predicted CG lightning days is high (median r=0.87), demonstrating that locally trained CNNs realistically capture year-to-year variation in CG lightning activity across the WUS. We then use layer-wise relevance propagation (LRP) to investigate the relevance of predictor variables to successful CG lightning prediction in each grid cell. Using maximum LRP values, our results show that two thermodynamic variables-ratio of surface moist static energy to free-tropospheric saturation moist static energy, and the 700-500hPa lapse rate-are the most relevant CG lightning predictors for 93%-96% of CNNs depending on the LRP variant used. As lightning is not directly simulated by global climate models, these CNNs could be used to parameterize CG lightning in climate models to assess changes in future CG lightning occurrence with projected climate change. Understanding changes in CG lightning risk and consequently lightning-caused wildfire risk across the WUS could inform fire management, planning, and disaster preparedness.

Comparison of Manual vs Artificial Intelligence-Based Muscle MRI Segmentation for Evaluating Disease Progression in Patients With CMT1A.

Intramuscular fat fraction (FF), assessed with quantitative MRI (qMRI), has emerged as one of the few responsive outcome measures in CMT1A patients. The main limitation for its use in future therapeutic trials is the time required for the manual segmentation of individual muscles. This study aimed to evaluate the accuracy and responsiveness of a fully automatic artificial intelligence (AI)-based segmentation pipeline to assess disease progression in a cohort of CMT1A patients over 1 year. Twenty CMT1A patients were included in this observational, prospective, longitudinal study. FF was measured twice a year apart using qMRI in the lower limbs. Individual muscle segmentation was performed fully automatically using a trained convolutional neural network with or without human quality check (QC). The corresponding results were compared with those obtained by fully manual (FM) segmentation using the Dice similarity coefficient (DSC). FF progression and its standardized response mean (SRM) were also computed in individual muscles over the single central slice and a 3D volume to define the most sensitive region of interest. AI-based segmentation showed excellent DSC values (>0.90). Significant global FF progression was observed at thigh (+0.71% ± 1.28%; p = 0.016) and leg (+1.73% ± 2.88%, p = 0.007) levels, similarly to that calculated using the FM technique (p = 0.363 and p = 0.634). FF progression of each individual muscle was comparable when computed from either the central slice or the 3D volume. The best SRM value (0.70) was obtained for the FF progression computed using the AI-based technique with human QC in the 3D volume at the leg level. The time required for fully automatic segmentation using AI with a QC was 10 hours for the entire data set compared with 90 hours for the FM. qMRI combined with AI-based segmentation can be considered as a process ready for assessing longitudinal FF changes in CMT1A patients. Given the slow FF progression at a thigh level and the large heterogeneity between muscles and individuals, FF should be quantified from a 3D volume at the leg level for longitudinal analyses. A QC performed after the AI-based segmentation is still advised given the increased SRM value.

GCNTrain+: A Versatile and Efficient Accelerator for Graph Convolutional Neural Network Training

Recently, graph convolutional networks (GCNs) have gained wide attention due to their ability to capture node relationships in graphs. One problem appears when full-batch GCN is trained on large graph datasets, where the computational and memory requirements are unacceptable. To address this issue, mini-batch GCN training is introduced to improve the scalability of GCN training for large datasets by sampling and training only a subset of the graph in each batch. Although several acceleration techniques have been designed for boosting the efficiency of full-batch GCN, they lack attention to mini-batch GCN, which differs from full-batch GCN in terms of the sampled dynamic graph structures. Based on our previous work GCNTrain [28], which was originally excogitated for accelerating full-batch GCN training, we devise GCNTrain+—a universal accelerator to tackle the performance bottlenecks associated with both full-batch and mini-batch GCN training. GCNTrain+ is equipped with two engines to optimize computation and memory access in GCN training, respectively. To reduce the computation overhead, we propose to dynamically reconfigure the computation order based on the varying data dimensions involved in each training batch. Moreover, we build a unified computation engine to perform the sparse-dense matrix multiplications (SpDM) and sparse-sparse matrix multiplications (SpSpM) discovered in GCN training uniformly. To alleviate the memory burden, we devise a two-phased dynamic clustering mechanism to capture data locality as well as customized hardware to reduce the clustering overhead. We evaluate GCNTrain+ on seven datasets, and the result shows that GCNTrain+ achieves 136.0 ×, 52.6 ×, 2.2 ×, and 1.5 × speedup over CPU, GPU, GCNAX, and GCNTrain in full-batch GCN training. Additionally, GCNTrain+ outperforms them with speedups of 131.6 ×, 67.1 ×, 4.4 ×, and 1.5 × in mini-batch GCN training.

Consensus-based iterative meta-pseudo-labeling for deep semi-supervised learning

A known issue that hinders the development of deep learning models is the need for accurate annotation of a large quantity of samples – a time-consuming, labor-intensive, and error-prone task. This limitation is particularly critical in areas where data annotation requires expert knowledge. Semi-supervised learning methods, such as pseudo-labeling, can alleviate the problem by capitalizing on both limited labeled and plentiful unlabeled data; nonetheless, state-of-the-art methods often require pre-trained encoders and validation sets to deliver effective solutions. Herein, we introduce a teacher-student-based iterative meta-pseudo-labeling approach, named consensus Deep Feature Annotation (cons-DeepFA), that enables the training of custom Convolutional Neural Networks (CNNs) from small quantities of labeled samples without reliance on pre-trained encoders and validation sets. cons-DeepFA explores Feature Learning from Image Markers (FLIM) to initialize the filters of a target CNN (student) from minimal data annotation – i.e., user-drawn markers on discriminative regions of a few selected images per class. During each of a few iterations, the latent space of the student's last dense layer is non-linearly projected onto a two-dimensional space for downstream label propagation via an optimum-connectivity-based approach (teacher); afterward, the student is re-trained using pseudo-labeled samples selected by the proposed consensus mechanism, which jointly improves the latent space, its projection, and the student's generalization ability as iterations progress. This strategy was recently introduced with pre-trained encoders by selecting the most confident pseudo-labeled samples to re-train the student. While building on previous methods, cons-DeepFA presents two key contributions. It (i) incorporates FLIM to enable training a custom CNN from scratch with faster convergence, improving its generalization ability, and (ii) introduces a consensus-based procedure over multiple iterations that selects more accurately pseudo-labeled samples for re-training the CNN. Lastly, cons-DeepFA is evaluated on five challenging biological image datasets, demonstrating its effectiveness and competitiveness when compared to seven state-of-the-art methods from four semi-supervised learning paradigms.

Preparing for downstream tasks in AI for dental radiology: a baseline performance comparison of deep learning models.

To compare the performance of the convolutional neural network (CNN) with the vision transformer (ViT) and the gated multilayer perceptron (gMLP) in the classification of radiographic images of dental structures. Retrospectively collected 2-dimensional images derived from cone beam computed tomographic volumes were used to train CNN, ViT and gMLP architectures as classifiers for 4 different cases. Cases selected for training the architectures were the classification of the radiographic appearance of maxillary sinuses, maxillary and mandibular incisors, presence or absence of the mental foramen and the positional relationship of the mandibular third molar to the inferior alveolar nerve canal. The performance metrics (sensitivity, specificity, precision, accuracy and f1-score) and area under curve (AUC) - receiver operating characteristic and precision-recall curves were calculated. The ViT with an accuracy of 0.74-0.98, performed on par with the CNN model (accuracy 0.71-0.99) in all tasks. The gMLP displayed marginally lower performance (accuracy 0.65-0.98) as compared to the CNN and ViT. For certain tasks, the ViT outperformed the CNN. The AUCs ranged from 0.77-1.00 (CNN), 0.80-1.00 (ViT) and 0.73-1.00 (gMLP) for all of the 4 cases. The difference in performance of the ViT, gMLP and the CNN (the current state-of-the-art) was significant in certain tasks. This difference in model performance for various tasks proves that capabilities of different architectures may be leveraged. The vision transformer, followed by the gated multilayer perceptron are deep learning models that exhibit comparable performance with the convolutional neural network in the classification of dental radiographic images.

Методи оптимізації та трансферне навчання для покращення виявлення пухлин мозку на основі МРТ-зображень

This research analyzes methods for optimizing the training of convolutional neural networks, the features of using small datasets, and the use of Transfer Learning techniques to improve image detection accuracy. The results of training these models were compared with each other on a test dataset for detecting brain tumor in CT images.

A dynamic strain prediction method for malfunction of sensors in buildings subjected to seismic loads using CWT and CNN

Structural health monitoring (SHM) is a technique used to evaluate the safety of a structure based on the structural response. In the SHM, sensors may fail to measure the structural response, or data loss can occur during the data transmission and reception. In this regard, this paper proposes a method of predicting the strain response when a sensor installed in a structural member suffers a malfunction. To this end, a convolutional neural network (CNN) is introduced to predict the strain response. Time-history strain data for seismic loads measured in adjacent structural members are set as the input layer of the CNN, and continuous wavelet transform (CWT) results are additionally set in the input layer of the CNN to reflect the dynamic structural behavior during earthquakes in the CNN predictions. Time-history strain data for a member, the target of prediction, are set as the output layer of the CNN. Then, the trained CNN uses the time-history strain data of adjacent members with no sensor failure and the CWT result of the data to predict the strain response of a structural member with sensor failure. A numerical study on the seismic load of a single-span three-story steel moment frame and an experimental study on a single-span three-story reinforced concrete frame were carried out to verify the validity of the proposed method. This study presents a method of applying CWT data to the input layer of the CNN and discusses the effectiveness of CWT data for predicting the nonlinear response of a structural member with damaged sensors.

Data augmentation via warping transforms for modeling natural variability in the corneal endothelium enhances semi-supervised segmentation.

Image segmentation of the corneal endothelium with deep convolutional neural networks (CNN) is challenging due to the scarcity of expert-annotated data. This work proposes a data augmentation technique via warping to enhance the performance of semi-supervised training of CNNs for accurate segmentation. We use a unique augmentation process for images and masks involving keypoint extraction, Delaunay triangulation, local affine transformations, and mask refinement. This approach accurately captures the natural variability of the corneal endothelium, enriching the dataset with realistic and diverse images. The proposed method achieved an increase in the mean intersection over union (mIoU) and Dice coefficient (DC) metrics of 17.2% and 4.8% respectively, for the segmentation task in corneal endothelial images on multiple CNN architectures. Our data augmentation strategy successfully models the natural variability in corneal endothelial images, thereby enhancing the performance and generalization capabilities of semi-supervised CNNs in medical image cell segmentation tasks.

Comprehensive Analysis of the Impact of Learning Rate and Dropout Rate on the Performance of Convolutional Neural Networks on the CIFAR-10 Dataset

Abstract. This study investigates the impact of two crucial hyperparameters, learning rate and dropout rate, on the performance of Convolutional Neural Networks (CNNs) on the CIFAR-10 dataset. Learning rate controls how quickly a model adapts to new data, while dropout rate helps prevent overfitting by randomly omitting neurons during training. These parameters are critical because they directly influence model convergence speed, accuracy, and generalization ability. By experimenting with various combinations of learning rates (0.01, 0.001, 0.0001) and dropout rates (0.3, 0.5, 0.7), we analyze the model's performance in terms of accuracy and overfitting. Our results suggest that a learning rate of 0.001 combined with a dropout rate of 0.5 yields the best balance between learning efficiency and generalization. This study offers important insights for optimizing hyperparameters in CNN training.

Removing atmospheric noise from InSAR interferograms in mountainous regions with a convolutional neural network

Atmospheric noise in interferometric synthetic aperture radar (InSAR)-derived estimates of surface deformation often obscures real displacement signals, especially in mountainous regions. As climate change disproportionately impacts the mountain cryosphere, a reliable technique for atmospheric correction in high-relief terrain is increasingly important. We developed and implemented a statistical machine learning atmospheric correction approach that relies on the differing spatial and topographic characteristics of slow-moving periglacial features and atmospheric noise. Our correction is applied at the native spatial and temporal resolution of the InSAR data, does not require external atmospheric reanalysis data, and can correct both stratified and turbulent atmospheric noise.Using Sentinel-1 data from 2017 to 2022, we trained a convolutional neural network (CNN) on observed atmospheric noise from 330 short-baseline interferograms and observed displacement signals from time series inversion of 1322 interferograms. We applied our trained CNN to correct 251 additional interferograms over an out-of-region application area, which were inverted to create displacement time series. We used the Rocky Mountains in New Mexico, Colorado, and Wyoming as our training, validation, testing, and application areas. When applied to our testing dataset, our correction offered performance improvements of 131%, 208%, and 68% in structural similarity index measure over corrections using atmospheric reanalysis data, phase correlation with topography, and high-pass filtering, respectively. The CNN-corrected time series reveals previously obscured kinematic behavior of rock glaciers and other features in the application dataset. Our flexible, robust approach can be used to correct arbitrary InSAR data to analyze subtle surface deformation signals for a range of science and engineering applications.

PIPET: A Pipeline to Generate PET Phantom Datasets for Reconstruction Based on Convolutional Neural Network Training

There has been a strong interest in using neural networks to solve several tasks in PET medical imaging. One of the main problems faced when using neural networks is the quality, quantity, and availability of data to train the algorithms. In order to address this issue, we have developed a pipeline that enables the generation of voxelized synthetic PET phantoms, simulates the acquisition of a PET scan, and reconstructs the image from the simulated data. In order to achieve these results, several pieces of software are used in the different steps of the pipeline. This pipeline solves the problem of generating diverse PET datasets and images of high quality for different types of phantoms and configurations. The data obtained from this pipeline can be used to train convolutional neural networks for PET reconstruction.

Development of an automated artificial intelligence-based system for urogenital schistosomiasis diagnosis using digital image analysis techniques and a robotized microscope.

Urogenital schistosomiasis is considered a Neglected Tropical Disease (NTD) by the World Health Organization (WHO). It is estimated to affect 150 million people worldwide, with a high relevance in resource-poor settings of the African continent. The gold-standard diagnosis is still direct observation of Schistosoma haematobium eggs in urine samples by optical microscopy. Novel diagnostic techniques based on digital image analysis by Artificial Intelligence (AI) tools are a suitable alternative for schistosomiasis diagnosis. Digital images of 24 urine sediment samples were acquired in non-endemic settings. S. haematobium eggs were manually labeled in digital images by laboratory professionals and used for training YOLOv5 and YOLOv8 models, which would achieve automatic detection and localization of the eggs. Urine sediment images were also employed to perform binary classification of images to detect erythrocytes/leukocytes with the MobileNetv3Large, EfficientNetv2, and NasNetLarge models. A robotized microscope system was employed to automatically move the slide through the X-Y axis and to auto-focus the sample. A total number of 1189 labels were annotated in 1017 digital images from urine sediment samples. YOLOv5x training demonstrated a 99.3% precision, 99.4% recall, 99.3% F-score, and 99.4% mAP0.5 for S. haematobium detection. NasNetLarge has an 85.6% accuracy for erythrocyte/leukocyte detection with the test dataset. Convolutional neural network training and comparison demonstrated that YOLOv5x for the detection of eggs and NasNetLarge for the binary image classification to detect erythrocytes/leukocytes were the best options for our digital image database. The development of low-cost novel diagnostic techniques based on the detection and identification of S. haematobium eggs in urine by AI tools would be a suitable alternative to conventional microscopy in non-endemic settings. This technical proof-of-principle study allows laying the basis for improving the system, and optimizing its implementation in the laboratories.

Multilevel attention mechanism for motion fatigue recognition based on sEMG and ACC signal fusion.

This study aims to develop a cost-effective and reliable motion monitoring device capable of comprehensive fatigue analysis. It achieves this objective by integrating surface electromyography (sEMG) and accelerometer (ACC) signals through a feature fusion strategy. The study introduces a multi-level attention mechanism for classification, leveraging convolutional neural networks (CNNs). The preprocessing phase involves a local feature attention mechanism that enhances local waveform features using the amplitude envelope. A dual-scale attention mechanism, operating at both channel and neuron levels, is employed to enhance the model's learning from high-dimensional fused data, improving feature extraction and generalization. The local feature attention mechanism significantly improves the model's classification accuracy and convergence, as demonstrated in ablation experiments. The model, optimized with multi-level attention mechanisms, excels in accuracy and generalization, particularly in handling data with pseudo-artifacts. Computational analysis indicates that the proposed optimization algorithm has minimal impact on CNN's training and testing times. The study achieves recognition accuracies of 92.52%, 92.38%, and 92.30%, as well as F1-scores of 91.92%, 92.13%, and 92.29% for the three fatigue states, affirming its reliability. This research provides technical support for the development of affordable and dependable wearable motion monitoring devices.

Gas Pipeline Leak Detection by Integrating Dynamic Modeling and Machine Learning Under the Transient State

This study focused on developing machine learning models to detect leak size and location in transient state conditions. The model was designed for an onshore methane–hydrogen blending gas pipeline in Canada. Base case simulations revealed significant effects on mass flow and pressure due to leaks, with the system taking approximately 6 h to reach a steady state from transient conditions. This made it essential to analyze the flow characteristics during the transient state. Trend data from the pipeline’s inlet and outlet were examined, considering the leak size and location. To better represent the data over time, a method was used to create two-dimensional images, which were then fed into a CNN (convolutional neural network) for training. The model’s accuracy was assessed using classification accuracy and a confusion matrix. By refining the data acquisition process and implementing targeted screening procedures, the model’s classification accuracy increased to over 80%. In conclusion, this study demonstrates that machine learning can enable rapid and accurate leak detection in transient state conditions. The findings are expected to complement existing leak detection methods and support operators in making faster, more informed decisions.

Automatic Segmentation of Quadriceps Femoris Cross-Sectional Area in Ultrasound Images: Development and Validation of Convolutional Neural Networks in People With Anterior Cruciate Ligament Injury and Surgery

ObjectiveDeep learning approaches such as DeepACSA enable automated segmentation of muscle ultrasound cross-sectional area (CSA). Although they provide fast and accurate results, most are developed using data from healthy populations. The changes in muscle size and quality following anterior cruciate ligament (ACL) injury challenges the validity of these automated approaches in the ACL population. Quadriceps muscle CSA is an important outcome following ACL injury; therefore, our aim was to validate DeepACSA, a convolutional neural network (CNN) approach for ACL injury. MethodsQuadriceps panoramic CSA ultrasound images (vastus lateralis [VL] n = 430, rectus femoris [RF] n = 349, and vastus medialis [VM] n = 723) from 124 participants with an ACL injury (age 22.8 ± 7.9 y, 61 females) were used to train CNN models. For VL and RF, combined models included extra images from healthy participants (n = 153, age 38.2, range 13–78) that the DeepACSA was developed from. All models were tested on unseen external validation images (n = 100) from ACL-injured participants. Model predicted CSA results were compared to manual segmentation results. ResultsAll models showed good comparability (ICC > 0.81, < 14.1% standard error of measurement, mean differences of <1.56 cm2) to manual segmentation. Removal of the erroneous predictions resulted in excellent comparability (ICC > 0.94, < 7.40% standard error of measurement, mean differences of <0.57 cm2). Erroneous predictions were 17% for combined VL, 11% for combined RF, and 20% for ACL-only VM models. ConclusionThe new CNN models provided can be used in ACL-injured populations to measure CSA of VL, RF, and VM muscles automatically. The models yield high comparability to manual segmentation results and reduce the burden of manual segmentation.

Artificial intelligence-based automated determination in breast and colon cancer and distinction between atypical and typical mitosis using a cloud-based platform.

Artificial intelligence (AI) technology in pathology has been utilized in many areas and requires supervised machine learning. Notably, the annotations that define the ground truth for the identification of different confusing process pathologies, vary from study to study. In this study, we present our findings in the detection of invasive breast cancer for the IHC/ISH assessment system, along with the automated analysis of each tissue layer, cancer type, etc. in colorectal specimens. Additionally, models for the detection of atypical and typical mitosis in several organs were developed using existing whole-slide image (WSI) sets from other AI projects. All H&E slides were scanned by different scanners with a resolution of 0.12-0.50μm/pixel, and then uploaded to a cloud-based AI platform. Convolutional neural networks (CNN) training sets consisted of invasive carcinoma, atypical and typical mitosis, and colonic tissue elements (mucosa-epithelium, lamina propria, muscularis mucosa, submucosa, muscularis propria, subserosa, vessels, and lymph nodes). In total, 59 WSIs from 59 breast cases, 217 WSIs from 54 colon cases, and 28 WSIs from 23 different types of tumor cases with relatively higher amounts of mitosis were annotated for the training. The harmonic average of precision and sensitivity was scored as F1 by AI. The final AI models of the Breast Project showed an F1 score of 94.49% for Invasive carcinoma. The mitosis project showed F1 scores of 80.18%, 97.40%, and 97.68% for mitosis, atypical, and typical mitosis layers, respectively. Overall F1 scores for the current results of the colon project were 90.02% for invasive carcinoma, 94.81% for the submucosa layer, and 98.02% for vessels and lymph nodes. After the training and optimization of the AI models and validation of each model, external validators evaluated the results of the AI models via blind-reader tasks. The AI models developed in this study were able to identify tumor foci, distinguish in situ areas, define colonic layers, detect vessels and lymph nodes, and catch the difference between atypical and typical mitosis. All results were exported for integration into our in-house applications for breast cancer and AI model development for both whole-block and whole-slide image-based 3D imaging assessment.

Kidney Disease Classification

Kidney diseases, such as tumors, cysts, and stones, are serious health conditions that affect millions globally. Medical imaging, particularly Computed-Tomography (CT) scans, plays an important role in diagnosing these conditions. Although, manual analysis of CT images by radiologist can be tiresome, fallible, and affected by human variability. This paper introduces a deep learning approach that employs Convolutional Neural Networks (CNN) for the automated classification of kidney CT scan images into four distinct categories : normal, cysts, tumors, and stones. The system is designed to boost diagnostic accuracy, reduce human error, and expedite clinical decision making. Through data preprocessing, CNN architecture design, training, and evaluation, the proposed model achieved a classification accuracy of 92%. This study highlights the significance of CNNs in medical image analysis anddtheir implicit for real- time deployment in clinical surroundings. Additionally, we compare our CNN-based approach with other advanced architectures like Vision Transformers, demonstrating CNNs' robustness and efficiency in this domain.

A student trained convolutional neural network competing with a commercial AI software and experts in organ at risk segmentation

This retrospective, multi-centered study aimed to improve high-quality radiation treatment (RT) planning workflows by training and testing a Convolutional Neural Network (CNN) to perform auto segmentations of organs at risk (OAR) for prostate cancer (PCa) patients, specifically the bladder and rectum. The objective of this project was to develop a clinically applicable and robust artificial intelligence (AI) system to assist radiation oncologists in OAR segmentation. The CNN was trained using manual contours in CT-datasets from diagnostic 68Ga-PSMA-PET/CTs by a student, then validated (n = 30, PET/CTs) and tested (n = 16, planning CTs). Further segmentations were generated by a commercial artificial intelligence (cAI) software. The ground truth were manual contours from expert radiation oncologists. The performance was evaluated using the Dice-Sørensen Coefficient (DSC), visual analysis and a Turing test. The CNN yielded excellent results in both cohorts and OARs with a DSCmedian > 0.87, the cAI resulted in a DSC > 0.78. In the visual assessment, 67% (bladder) and 75% (rectum) of the segmentations were rated as acceptable for treatment planning. With a misclassification rate of 45.5% (bladder) and 51.1% (rectum), the CNN passed the Turing test. The metrics, visual assessment and the Turing test confirmed the clinical applicability and therefore the support in clinical routine.

Transfer learning classification of suspicious lesions on breast ultrasound: is there room to avoid biopsies of benign lesions?

BackgroundBreast cancer (BC) is the most common malignancy in women and the second cause of cancer death. In recent years, there has been a strong development in artificial intelligence (AI) applications in medical imaging for several tasks. Our aim was to evaluate the potential of transfer learning with convolutional neural networks (CNNs) in discriminating suspicious breast lesions on ultrasound images.MethodsTransfer learning performances of five different CNNs (Inception V3, Xception, Densenet121, VGG 16, and ResNet50) were evaluated on a public and on an institutional dataset (526 and 392 images, respectively), customizing the top layers for the specific task. Institutional images were contoured by an expert radiologist and processed to feed the CNNs for training and testing. Postimaging biopsies were used as a reference standard for classification. The area under the receiver operating curve (AUROC) was used to assess diagnostic performance.ResultsNetworks performed very well on the public dataset (AUROC 0.938–0.996). The direct generalization to the institutional dataset resulted in lower performances (max AUROC 0.676); however, when tested on BI-RADS 3 and BI-RADS 5 only, results were improved (max AUROC 0.792). Good results were achieved on the institutional dataset (AUROC 0.759–0.818) and, when selecting a threshold of 2% for classification, a sensitivity of 0.983 was obtained for three of five CNNs, with the potential to spare biopsy in 15.3%–18.6% of patients.ConclusionIn conclusion, transfer learning with CNNs may achieve high sensitivity and might be used as a support tool in managing suspicious breast lesions on ultrasound images.Relevance statementTransfer learning is a powerful technique to exploit the performances of well-trained CNNs for image classification. In a clinical scenario, it might be useful for the management of suspicious breast lesions on breast ultrasound, potentially sparing biopsy in a non-negligible number of patients.Key PointsProperly trained CNNs with transfer learning are highly effective in differentiating benign and malignant lesions on breast ultrasound.Setting clinical thresholds increased sensitivity.CNNs might be useful as support tools in managing suspicious lesions on breast ultrasound.Graphical

Behavioral Cloning Strategies in Steering Angle Prediction: Applications in Mobile Robotics and Autonomous Driving

Artificial neural networks (ANNs) are artificial intelligence techniques that have made autonomous driving more efficient and accurate; however, autonomous driving faces ongoing challenges in the accuracy of decision making based on the analysis of the vehicle environment. A critical task of ANNs is steering angle prediction, which is essential for safe and effective navigation of mobile robots and autonomous vehicles. In this study, to optimize steering angle prediction, NVIDIA’s architecture was adapted and modified along with the implementation of the Swish activation function to train convolutional neural networks (CNNs) by behavioral cloning. The CNN used human driving data obtained from the UDACITY beta simulator and tests in real scenarios, achieving a significant improvement in the loss function during training, indicating a higher efficiency in replicating human driving behavior. The proposed neural network was validated through implementation on a differential drive mobile robot prototype, by means of a comparative analysis of trajectories in autonomous and manual driving modes. This work not only advances the accuracy of steering angle predictions but also provides valuable information for future research and applications in mobile robotics and autonomous driving. The performance results of the model trained with the proposed CNN show improved accuracy in various operational contexts.