DL Models Research Articles

Using Adversarial Images to Assess the Robustness of Deep Learning Models Trained on Diagnostic Images in Oncology.

PURPOSEDeep learning (DL) models have rapidly become a popular and cost-effective tool for image classification within oncology. A major limitation of DL models is their vulnerability to adversarial images, manipulated input images designed to cause misclassifications by DL models. The purpose of the study is to investigate the robustness of DL models trained on diagnostic images using adversarial images and explore the utility of an iterative adversarial training approach to improve the robustness of DL models against adversarial images.METHODSWe examined the impact of adversarial images on the classification accuracies of DL models trained to classify cancerous lesions across three common oncologic imaging modalities. The computed tomography (CT) model was trained to classify malignant lung nodules. The mammogram model was trained to classify malignant breast lesions. The magnetic resonance imaging (MRI) model was trained to classify brain metastases.RESULTSOncologic images showed instability to small pixel-level changes. A pixel-level perturbation of 0.004 (for pixels normalized to the range between 0 and 1) resulted in most oncologic images to be misclassified (CT 25.6%, mammogram 23.9%, and MRI 6.4% accuracy). Adversarial training improved the stability and robustness of DL models trained on oncologic images compared with naive models ([CT 67.7% v 26.9%], mammogram [63.4% vs 27.7%], and MRI [87.2% vs 24.3%]).CONCLUSIONDL models naively trained on oncologic images exhibited dramatic instability to small pixel-level changes resulting in substantial decreases in accuracy. Adversarial training techniques improved the stability and robustness of DL models to such pixel-level changes. Before clinical implementation, adversarial training should be considered to proposed DL models to improve overall performance and safety.

Design and implementation of information extraction system for scientific literature using fine-tuned deep learning models

This paper presents an overview of a quality scoring system that utilizes pre-trained deep neural network models. Two types of DL models, a classification and extractive question answering (EQA) models are used to implement components of the system. The abstracts of the scientific literature are classified into two groups, in-vivo and in-vitro, and a question and answering model architecture is constructed for extracting the following types of information (animal type, the number of animals, exposure dose, and signal frequency). The Bidirectional Encoder Representations of Transformers (BERT) model pre-trained with a large text corpus is used as our baseline model for classification and EQA tasks. The models are fine-tuned with 455 EMF-related research papers. In our experiments, the fine-tuned model showed improved performance on EQA tasks for the four-categories of questions compared to the baseline, and it also showed improvements on similar questions that were not used in training. This suggests the importance of retraining of deep learning model specifically in some areas requiring domain expertise such as scientific papers. However, additional research is needed on some implementation issues, in such cases where there are still multiple answers, or where there is no answer given in a context.

Abstract ES7-3:/AI

Abstract The tools of AI offer promise to uncover and amplify data, previously hidden in images of the human body, and expand the field of radiomics to more accurately predict breast cancer risk across diverse patient populations. Although deep learning products trained to predict mammographic breast density and risk of breast cancer are now available, their performance in routine clinical settings is largely unknown. Since the creation of the Gail model in 1989, risk models have supported risk-adjusted screening and prevention, and their continued evolution has been a central pillar of breast cancer research. Mammographic breast density is now being incorporated into well-established clinical risk models, which are used to determine eligibility for supplemental imaging and other services for patients at increased risk. However, mammographic breast density assessment is subjective and varies widely between and within radiologists. For example, Sprague et al demonstrated considerable variability of qualitative Breast Imaging Reporting and Data System (BI-RADS) density assessments with a range of 6% to 85% of mammograms being assessed as heterogeneously or extremely dense across 83 radiologists. To limit variability and subjectivity, various automated breast density assessment methods were commercially developed, but clinical evaluation has yielded mixed results. Density is only one limited feature of any woman’s mammogram. Deep learning models can operate over the full resolution of mammogram images to assess a patient’s future breast cancer risk. Rather than manually identifying discriminative image patterns, machine learning models can discover these patterns directly from the data. Specifically, models are trained with full resolution mammograms and the outcome of interest, namely whether the patient developed breast cancer within five years from the date of the examination. Our recent work demonstrates that application of novel artificial intelligence applications to imaging data can significantly improve breast density assessment and breast cancer risk prediction. In addition, unlike traditional models, our DL models perform equally well across varied races, ages, and family histories and we have built a clinical platform which is currently in use to support implementation of our density and risk models into routine clinical care. From this clinical platform, we found that use of a deep learning model more accurately predicts mammographic breast density, reducing human variation and “overcalling” of breast density, and can help healthcare systems more appropriately utilize limited supplemental screening resources as well as provide patients with more accurate information regarding their breast density. We have also found a DL breast cancer risk model, generated in routine clinical mammography screening programs, can predict future breast cancer risk and can more accurately identify women who will most benefit from screening mammography. This DL score can be used in diverse clinical settings that currently depend on traditional risk scores, which are at best modest in performance. DL risk scores based on the mammogram alone can inform shared decision making with patients and their healthcare providers regarding more personalized risk reduction and early detection strategies. Citation Format: C Lehman, B Dontchos, LR Lamb./AI [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr ES7-3.

Trend decomposition aids forecasts of air particulate matter (PM2.5) assisted by machine and deep learning without recourse to exogenous data

A near-past, trend-attribute extraction technique is proposed for short-term hourly particulate matter (PM2.5) forecasting. Multiple attributes are extracted from the univariate PM2.5 time-series data for the prior 12 h to forecast PM2.5 for the next 12 h with the aid of machine/deep learning (ML/DL) algorithms. The method avoids the uncertainties of using a selection of exogeneous variables to forecast PM2.5 trends that tend to be highly volatile with erratically spaced peaks. Fifteen trend-decomposed attributes are extracted for each hourly data record of a 2016 to 2020 dataset for Nottingham Centre (a city in the United Kingdom). The 41979 data records are evaluated by sixteen ML/DL forecasting models. Each model is trained and validated with two datasets (2016–2018 and 2016–2019) that are then used to forecast 5 h-ahead periods (t0, t+1, t+3, t+6 and t+12) for 2019 and 2020, respectively. ML models outperform DL models in forecasting PM2.5 for 2019 and 2020, because the DL models are adversely impacted by suboptimal convergence and/or overfitting. The best performing models are multiple linear regression (MLR) and support vector regression (SVR) that outperform traditional timeseries autoregressive models which estimate a parametric autocorrelation parameter, substantially so for time periods beyond t+1. The transparent MLR model reveals that all the trend-attribute variables considered influence the forecasts of the trained models, but their relative contribution varies according to the time period being forecast. Enabling the models to allocate different weights to each variable for different time period forecasts adds a useful dimension of flexibility to this method compared to autoregressive models.

Optimal Deep-Learning-Enabled Intelligent Decision Support System for SARS-CoV-2 Classification.

Intelligent decision support systems (IDSS) for complex healthcare applications aim to examine a large quantity of complex healthcare data to assist doctors, researchers, pathologists, and other healthcare professionals. A decision support system (DSS) is an intelligent system that provides improved assistance in various stages of health-related disease diagnosis. At the same time, the SARS-CoV-2 infection that causes COVID-19 disease has spread globally from the beginning of 2020. Several research works reported that the imaging pattern based on computed tomography (CT) can be utilized to detect SARS-CoV-2. Earlier identification and detection of the diseases is essential to offer adequate treatment and avoid the severity of the disease. With this motivation, this study develops an efficient deep-learning-based fusion model with swarm intelligence (EDLFM-SI) for SARS-CoV-2 identification. The proposed EDLFM-SI technique aims to detect and classify the SARS-CoV-2 infection or not. Also, the EDLFM-SI technique comprises various processes, namely, data augmentation, preprocessing, feature extraction, and classification. Moreover, a fusion of capsule network (CapsNet) and MobileNet based feature extractors are employed. Besides, a water strider algorithm (WSA) is applied to fine-tune the hyperparameters involved in the DL models. Finally, a cascaded neural network (CNN) classifier is applied for detecting the existence of SARS-CoV-2. In order to showcase the improved performance of the EDLFM-SI technique, a wide range of simulations take place on the COVID-19 CT data set and the SARS-CoV-2 CT scan data set. The simulation outcomes highlighted the supremacy of the EDLFM-SI technique over the recent approaches.

Performance Analysis of Channel Estimation for Massive MIMO Communication Using DL-Based Fully Connected Neural Network (DL-FCNN) Architecture

The latest research for applying deep learning in wireless communications gives several opportunities to reduce complex signal processing. The channel estimation is important to study the nature of the varying channel and to calculate channel state information (CSI) value which is utilized at the receiver to nullify the interference which occurs during multipath transmission. In the current article, considering the massive Multiple Input Multiple Output (MIMO) channel model, a DL approach is developed with a fully connected neural network (NN) architecture which is used to estimate the channel with minimum error. The proposed DL architecture uses an openly available channel dataset. Further, using generated pilot symbols of lengths 2 and 4, the performance of DL-based Fully connected NN (DL-FCNN) is analyzed to estimate the channel in uplink massive MIMO communication. The obtained results demonstrate that the channel estimation performance was calculated in terms of normalized mean square error((NMSE) for different values of SNR added at receiver base station (BS) to the signals over the range of BS antennas. Also, the channel estimation error over a large number of BS antennas for massive MIMO scenarios is observed, and it is observed that the NMSE reduces with a greater number of antennas. Hence, it can be inferred that the DL models will be the future for most physical layer signal processing techniques such as channel estimation, modulation detection, etc. within massive MIMO networks.

Sentimental analysis on user’s reviews using BERT

In the current E-Commerce and M−Commerce evolved Internet world, people are expressing their sentiments on the product or service in the form of reviews and ratings. Sentiment analysis finds a contextual inference in the user sentiment which helps the E-Commerce sector and other end-users to realize the opinion of the product or service. People give their opinion in a variety of formats like reviews, blogs, news, or comments. Sentiment Analysis is used quick gaining insights using large volumes of text data which can be helpful in improving the quality of service and can make companies to earn huge profits. In sentimental analysis, modeling of sentimental relations such as word negation and word intensification are great challenge. In the proposed work, Bidirectional Encoder Representations from Transformers (BERT) is constructed for classifying user sentiments on IMDB Movie Review and Amazon Fine Food Review datasets. Word negations and intensifications are also considered for deriving the sentiment of reviews. The experimental results shows that the BERT can capture the sentimental relations effectively than basic ML, DL models and other recent literatures.

A Framework for implementing an ML or DL model to improve Intrusion Detection Systems (IDS) in the NTMA context, with an example on the dataset (CSE-CIC-IDS2018)

The objective of this work is to present a framework to be followed to model, test, validate and implement a DL model for anomaly, abuse, malware or botnet detection, with the aim of implementing or improving an Intrusion Detection System (IDS) within the NTMA framework, by means of new machine learning and deep learning techniques, which addresses reliability and processing speed considerations. The said process will be used to perform studies on ML and DL models used for cybersecurity in isolation and in combination to extract conclusions, which can help in the improvement of intrusion detection systems using massive data collection techniques used in Big-Data. The example discussed in this work implemented part of our framework by applying the CNN algorithm on the CSE-CIC-IDS2018 dataset. The results are encouraging for the use of ML in IDS, with an efficiency that exceeds 92% after 30 iterations. Thus, this model remains to be improved and tested on real networks.

Plant phenotyping with limited annotation: Doing more with less

AbstractDeep learning (DL) methods have transformed the way we extract plant traits—both under laboratory as well as field conditions. Evidence suggests that “well‐trained” DL models can significantly simplify and accelerate trait extraction as well as expand the suite of extractable traits. Training a DL model typically requires the availability of copious amounts of annotated data; however, creating large‐scale annotated dataset requires nontrivial efforts, time, and resources. This limitation has become a major bottleneck in deploying DL tools in practice. Self‐supervised learning (SSL) methods give exciting solution to this problem, as these methods use unlabeled data to produce pretrained models for subsequent fine‐tuning on labeled data and have demonstrated superior transfer learning performance on down‐stream classification tasks. We investigated the application of SSL methods for plant stress classification using few labels. We select a plant stress classification problem to test the effectiveness of SSL, as it is a fundamentally challenging problem due to (a) disease classification which depends on the abnormalities in a small number of pixels, (b) high data imbalance across different classes, and (c) fewer annotated and available plant stress images than in other domains. We compared seven SSL approaches spanning four broad classes of SSL methods on soybean [Glycine max L. (Merr.)] plant stress dataset and report that pretraining on unlabeled plant stress images significantly outperforms transfer learning methods using random initialization for plant stress classification. In summary, SSL‐based model initialization and data curation improves annotation efficiency for plant stress classification tasks and will circumvent data annotation challenges associated with DL methods.

Evolutionary Intelligence and Deep Learning Enabled Diabetic Retinopathy Classification Model

Diabetic Retinopathy (DR) has become a widespread illness among diabetics across the globe. Retinal fundus images are generally used by physicians to detect and classify the stages of DR. Since manual examination of DR images is a time-consuming process with the risks of biased results, automated tools using Artificial Intelligence (AI) to diagnose the disease have become essential. In this view, the current study develops an Optimal Deep Learning-enabled Fusion-based Diabetic Retinopathy Detection and Classification (ODL-FDRDC) technique. The intention of the proposed ODL-FDRDC technique is to identify DR and categorize its different grades using retinal fundus images. In addition, ODL-FDRDC technique involves region growing segmentation technique to determine the infected regions. Moreover, the fusion of two DL models namely, CapsNet and MobileNet is used for feature extraction. Further, the hyperparameter tuning of these models is also performed via Coyote Optimization Algorithm (COA). Gated Recurrent Unit (GRU) is also utilized to identify DR. The experimental results of the analysis, accomplished by ODL-FDRDC technique against benchmark DR dataset, established the supremacy of the technique over existing methodologies under different measures.

Silver nano-particles' size prediction using deep learning model with green method

When it comes to casting technologies and other production processes such as welding, DL models minimise the need for costly experimental inquiry. One of the most important criteria in nanotechnology is the measurement of particle size. They've received a lot of interest for their chemical, physical, and clinical uses because of their outstanding qualities. By adopting a green chemical reduction process, the nano-silver crystals are generated in the biopolymer interface without any agglomeration. Nano-metal production relies on precise control of particle size, and this approach provides that. For the prediction size of Ag-NPs via the green approach, a hybrid ideal of Long-Short Term Memory (LSTM) and Convolutional Neural Network (CNN) and model is presented in this research. Temperature, starch stabiliser, and AgNO3 concentration are among the experimental variables fed into the CNN-LSTM model. The size of a nano particle is the output of the CNN-LSTM model. Making three subsets of the experimental data for training, validation, and testing was one of the initial steps in creating an adapting model. The silver nanoparticles were identified using X-ray diffraction and transmission electron microscopy. Ag-NPs' crystallographic airplanes were found to be face-centered cubic, according to XRD analysis (fcc).When comparing with other existing classifiers, DNN achieved high performance on precision, i.e. 89.87%, and this is due to the presence of simple structure. The CNN and auto encoder achieved nearly 82% and 80% of accuracy, where LSTM achieved 77.25% of accuracy and proposed CNN-LSTM achieved 83.50% of recall.

Wasserstein GAN-gradient penalty with deep transfer learning based alzheimer disease classification on 3D MRI scans

There has been growing interest in using neuroimaging data, such as MRI scans, for the detection of Alzheimer's Disease (AD). Computer vision and deep learning models have shown promise in developing effective Computer-Aided Diagnosis (CAD) models for AD detection and classification. However, many existing models struggle due to their reliance on large training datasets and effective hyper parameter tuning strategies. To address these issues, transfer learning is often used to adjust the final fully connected layers of pre-trained DL models for use with smaller datasets. This paper proposes a new AD classification model based on a combination of Wasserstein GAN-Gradient Penalty (WGANGP) and Deep Transfer Learning (DTL) techniques, aimed at achieving accurate identification and classification of AD on 3D MRI scans. The WGANGP technique is used to increase the size of the dataset, and the model utilizes image enhancement and 3D Spatial Fuzzy C-means (3DS-FCM) techniques for image segmentation. Additionally, feature extraction is performed using the Ant Lion Optimizer (ALO) with the Inception v3 model, while the Deep Belief Network (DBN) model is employed for AD classification. The experimental validation of the WGANGP-DTL model is conducted using a benchmark 3D MRI dataset, and the results show that the proposed model outperforms recent approaches in several aspects.

Deep Learning Inspired Feature Engineering for Classifying Tremor Severity

Bio-signals pattern recognition systems can be impacted by several factors with a potential to limit their associated performance and clinical translation. Among these factors, selecting the optimum feature extraction method, that can effectively exploit the interaction between the temporal and spatial information, is the most prominent. Despite the potential of DL models for extracting temporal, spatial, or temporal-spatial information, they are typically restricted by their needs for a large amount of training data. The deep wavelet scattering transform (WST) is a relatively recent advancement within the DL literature to replace expensive convolution neural networks models with computationally less demanding methods. However, while some studies have used WST to extract features from biological signals, it has not been widely investigated before for electromyogram (EMG) and electroencephalogram (EEG) signals feature extraction. To investigate the hypothesis of the usefulness of WST for processing EMG and EEG signals, this study used a tremor dataset collected by the authors from people with tremor disorders. Specifically, the proposed work achieved three goals: (a) to study the performance of extracting features from low-density EMG signals (8 channels), using the WST approach, (b) to study the effect of extracting the features from high-density EEG signals (33 channels), using WST and study its robustness against changing spatial and temporal aspects of classification accuracy, and (c) to classify tremor severity using the WST method and compare the results with other well-known feature extraction approaches. The classification error rates were significantly reduced (nearly 15%) compared with other known feature sets.

Insider Threat Detection Based on NLP Word Embedding and Machine Learning

The growth of edge computing, the Internet of Things (IoT), and cloud computing have been accompanied by new security issues evolving in the information security infrastructure. Recent studies suggest that the cost of insider attacks is higher than the external threats, making it an essential aspect of information security for organizations. Efficient insider threat detection requires state-of-the-art Artificial Intelligence models and utility. Although significant have been made to detect insider threats for more than a decade, there are many limitations, including a lack of real data, low accuracy, and a relatively low false alarm, which are major concerns needing further investigation. In this paper, an attempt to fulfill these gaps by detecting insider threats with the novelties of the present investigation first developed two deep learning hybrid LSTM models integrated with Google's Word2vec LSTM (Long Short-Term Memory) GLoVe (Global Vectors for Word Representation) LSTM. Secondly, the performance of two hybrid DL models was compared with the state-of-the-art ML models such as XGBoost, AdaBoost, RF (Random Forest), KNN (K-Nearest Neighbor) and LR (Logistics Regression). Thirdly, the present investigation bridges the gaps of using a real dataset, high accuracy, and significantly lower false alarm rate. It was found that ML-based models outperformed the DL-based ones. The results were evaluated based on earlier studies and deemed efficient at detecting insider threats using the real dataset.

Сравнительный анализ инструментов непрерывной онлайн-аутентификации и систем обнаружения аномалий для постоянного подтверждения личности пользователя

Problem statement: increasing the security of the system actualizes issues related to the confirmation of the user's identity by various methods. Known ways to increase security, continuous authentication by biometric parameters, such as the movement of a computer mouse, the speed of clicking mouse buttons is complex and does not give one hundred percent accuracy. The purpose of the work: is to study the existing solutions in the field of anomaly detection and continuous authentication systems. It is proposed to develop parameters for removing data from a computer mouse, requirements for creating a DataSet, the effectiveness of analysis and extraction of the main features from the raw data. Consider the DL models for CA and AD, allowing to verify the user, to make their qualitative comparison. Methods used: the study uses a comparative analysis of technologies. Such as differentiated machine learning (ML), including the decision tree classifier (DT), the k-nearest neighbor method (k-NN), the random forests algorithm (RF) and ultra-precise neural networks (CNN). Result: based on a comparative analysis of continuous online authentication tools and anomaly detection systems for permanent confirmation of the user's identity. As a result of the work, it was determined that the existing methods related to CA and AD describe mainly laboratory studies, and are in the nature of confirming hypotheses. The considered studies describe the parameters used to develop software for taking data about mouse actions. Studies show that the collection tools have little effect on the quality of the data obtained, provided that the data is preprocessed using algorithms for identifying key features. Most studies consider data sets obtained either from free sources or from groups of volunteers ranging from 10 to 60 people. It should be noted that the number exceeding 10 can be considered sufficient to determine a representative sample. It was determined that the greatest value on the result was given by signs describing the movement of the mouse between two locations of the screen. The DL models for CA and AD that allow user verification are considered, and their qualitative comparison is made. Practical significance: the conducted research makes it possible to judge the possibility of implementing a solution for continuous online authentication and anomaly detection systems for permanent confirmation of the user's identity within the framework of the RemoteTopology remote championship system developed earlier.

An attention-aware LSTM model for soil moisture and soil temperature prediction

Accurate prediction of soil moisture (SM) and soil temperature (ST) plays an important role in Earth system science, helping to forecast and understand ecosystem changes. They present great challenges because land-atmospheric interactions are complex and diverse in space and time. Although deep learning methods have excellent performance for land surface variables’ prediction such as SM and ST, they are often questioned due to their over-parameterized black-box nature and neglect of physical knowledge and interpretability. From this, we propose an attention-aware LSTM Model (ILSTM_Soil) by taking multi-feature attention, predictor attention and temporal attention into account. We first used LSTM to generate multi-feature vectors of all predictors, and then the three attention mechanisms were designed to summarize these feature vectors for SM and ST prediction. Experiment results for SM and ST prediction at the lead time of 1 and 7 days on ten FLUXNET sites suggest that the proposed ILSTM_Soil model outperforms Random Forest (RF), Support Vector Regression (SVR), Elastic-Net (ENET), original Long Short-Term Memory (LSTM), and attention LSTM (A-LSTM) models in most cases. The interpretation of attention weights verifies that the proposed model can capture physical knowledge over SM and ST. The code of ILSTM_Soil is made publicly available and we hope it can encourage researchers to develop effective DL models in land surface variables’ prediction conveniently.

Application of CycleGAN and transfer learning techniques for automated detection of COVID-19 using X-ray images.

Coronavirus (which is also known as COVID-19) is severely impacting the wellness and lives of many across the globe. There are several methods currently to detect and monitor the progress of the disease such as radiological image from patients’ chests, measuring the symptoms and applying polymerase chain reaction (RT-PCR) test. X-ray imaging is one of the popular techniques used to visualise the impact of the virus on the lungs. Although manual detection of this disease using radiology images is more popular, it can be time-consuming, and is prone to human errors. Hence, automated detection of lung pathologies due to COVID-19 utilising deep learning (Bowles et al.) techniques can assist with yielding accurate results for huge databases. Large volumes of data are needed to achieve generalizable DL models; however, there are very few public databases available for detecting COVID-19 disease pathologies automatically. Standard data augmentation method can be used to enhance the models’ generalizability. In this research, the Extensive COVID-19 X-ray and CT Chest Images Dataset has been used and generative adversarial network (GAN) coupled with trained, semi-supervised CycleGAN (SSA- CycleGAN) has been applied to augment the training dataset. Then a newly designed and finetuned Inception V3 transfer learning model has been developed to train the algorithm for detecting COVID-19 pandemic. The obtained results from the proposed Inception-CycleGAN model indicated Accuracy = 94.2%, Area under Curve = 92.2%, Mean Squared Error = 0.27, Mean Absolute Error = 0.16. The developed Inception-CycleGAN framework is ready to be tested with further COVID-19 X-Ray images of the chest.

Insights of an AI agent via analysis of prediction errors: a case study of fluence map prediction for radiation therapy planning

Purpose. We have previously reported an artificial intelligence (AI) agent that automatically generates intensity-modulated radiation therapy (IMRT) plans via fluence map prediction, by-passing inverse planning. This AI agent achieved clinically comparable quality for prostate cases, but its performance on head-and-neck patients leaves room for improvement. This study aims to collect insights of the deep-learning-based (DL-based) fluence map prediction model by systematically analyzing its prediction errors. Methods. From the modeling perspective, the DL model’s output is the fluence maps of IMRT plans. However, from the clinical planning perspective, the plan quality evaluation should be based on the clinical dosimetric criteria such as dose-volume histograms. To account for the complex and non-intuitive relationships between fluence map prediction errors and the corresponding dose distribution changes, we propose a novel error analysis approach that systematically examines plan dosimetric changes that are induced by varying amounts of fluence prediction errors. We investigated four decomposition modes of model prediction errors. The two spatial domain decompositions are based on fluence intensity and fluence gradient. The two frequency domain decompositions are based on Fourier-space banded frequency rings and Fourier-space truncated low-frequency disks. The decomposed error was analyzed for its impact on the resulting plans’ dosimetric metrics. The analysis was conducted on 15 test cases spared from the 200 training and 16 validation cases used to train the model. Results. Most planning target volume metrics were significantly correlated with most error decompositions. The Fourier space disk radii had the largest Spearman’s coefficients. The low-frequency region within a disk of ∼20% Fourier space contained most of errors that impact overall plan quality. Conclusions. This study demonstrates the feasibility of using fluence map prediction error analysis to understand the AI agent’s performance. Such insights will help fine-tune the DL models in architecture design and loss function selection.

Ensemble of Deep Learning-Based Multimodal Remote Sensing Image Classification Model on Unmanned Aerial Vehicle Networks

Recently, unmanned aerial vehicles (UAVs) have been used in several applications of environmental modeling and land use inventories. At the same time, the computer vision-based remote sensing image classification models are needed to monitor the modifications over time such as vegetation, inland water, bare soil or human infrastructure regardless of spectral, spatial, temporal, and radiometric resolutions. In this aspect, this paper proposes an ensemble of DL-based multimodal land cover classification (EDL-MMLCC) models using remote sensing images. The EDL-MMLCC technique aims to classify remote sensing images into the different cloud, shades, and land cover classes. Primarily, median filtering-based preprocessing and data augmentation techniques take place. In addition, an ensemble of DL models, namely VGG-19, Capsule Network (CapsNet), and MobileNet, is used for feature extraction. In addition, the training process of the DL models can be enhanced by the use of hosted cuckoo optimization (HCO) algorithm. Finally, the salp swarm algorithm (SSA) with regularized extreme learning machine (RELM) classifier is applied for land cover classification. The design of the HCO algorithm for hyperparameter optimization and SSA for parameter tuning of the RELM model helps to increase the classification outcome to a maximum level considerably. The proposed EDL-MMLCC technique is tested using an Amazon dataset from the Kaggle repository. The experimental results pointed out the promising performance of the EDL-MMLCC technique over the recent state of art approaches.

A Convolutional Autoencoder Topology for Classification in High-Dimensional Noisy Image Datasets.

Deep convolutional neural networks have shown remarkable performance in the image classification domain. However, Deep Learning models are vulnerable to noise and redundant information encapsulated into the high-dimensional raw input images, leading to unstable and unreliable predictions. Autoencoders constitute an unsupervised dimensionality reduction technique, proven to filter out noise and redundant information and create robust and stable feature representations. In this work, in order to resolve the problem of DL models’ vulnerability, we propose a convolutional autoencoder topological model for compressing and filtering out noise and redundant information from initial high dimensionality input images and then feeding this compressed output into convolutional neural networks. Our results reveal the efficiency of the proposed approach, leading to a significant performance improvement compared to Deep Learning models trained with the initial raw images.