DL-based Model Research Articles

Selective functionalization of hindered meta-C–H bond of o-alkylaryl ketones promoted by automation and deep learning

<h2>Summary</h2> Selective functionalization of the sterically hindered aromatic <i>meta</i>-C–H bond is unprecedented and remains to be a major challenge. Promoted by automation-based high-throughput experimentation (HTE) and deep learning (DL), a novel strategy to functionalize the hindered <i>meta</i>-C–H bond is disclosed. With carbon dioxide as a traceless director, a one-pot three-step protocol was developed to achieve selective arylation of <i>o</i>-alkylaryl ketones at the hindered <i>meta</i> position. This novel strategy involved photo-induced C–H carboxylation, carboxyl group-directed Pd-catalyzed C–H functionalization, and microwave-assisted decarboxylation. With HTE and DL, a broad scope of substrates was explored (1,032 reactions) and a DL-based model (CMPRY) for reaction yield prediction was established. Two independent tests with unseen <i>o</i>-alkylaryl ketones and/or potassium aryltrifluoroborates were used to evaluate the model. The model gave excellent performances in predicting unseen reactions; mean absolute errors in yield were only 6.6% and 8.4%, suggesting its potential in synthetic application.

An Interpretable Deep Learning Method for Power System Transient Stability Assessment via Tree Regularization

Deep learning (DL) techniques have shown promising performance for designing data-driven power system transient stability assessment (TSA) models. However, due to the deep structure of the DL, the resulting model is always a black-box and hard to explain, which hinders its practical adoption by the industry. This paper proposes an interpretable DL-based TSA model to balance the TSA accuracy and transparency. The proposed method combines the strong nonlinear modelling capability of a deep neural network and the interpretability of a decision tree (DT). Through regularizing DL-based model with the average DT path length in the training process, the proposed interpretable DL-based TSA method can visually explain the TSA decision-making process. Simulation results have shown that the proposed method can deliver highly accurate TSA results and interpretable TSA decision-making rules, which can be used for designing preventive control actions.

Systematic Clinical Evaluation of a Deep Learning Method for Medical Image Segmentation: Radiosurgery Application.

We systematically evaluate a Deep Learning model in a 3D medical image segmentation task. With our model, we address the flaws of manual segmentation: high inter-rater contouring variability and time consumption of the contouring process. The main extension over the existing evaluations is the careful and detailed analysis that could be further generalized on other medical image segmentation tasks. Firstly, we analyze the changes in the inter-rater detection agreement. We show that the model reduces the number of detection disagreements by [Formula: see text] [Formula: see text]. Secondly, we show that the model improves the inter-rater contouring agreement from [Formula: see text] to [Formula: see text] surface Dice Score [Formula: see text]. Thirdly, we show that the model accelerates the delineation process between [Formula: see text] and [Formula: see text] times [Formula: see text]. Finally, we design the setup of the clinical experiment to either exclude or estimate the evaluation biases; thus, preserving the significance of the results. Besides the clinical evaluation, we also share intuitions and practical ideas for building an efficient DL-based model for 3D medical image segmentation.

Deep Learning-Based Multi-Omics Integration Robustly Predicts Relapse in Prostate Cancer.

ObjectivePost-operative biochemical relapse (BCR) continues to occur in a significant percentage of patients with localized prostate cancer (PCa). Current stratification methods are not adequate to identify high-risk patients. The present study exploits the ability of deep learning (DL) algorithms using the H2O package to combine multi-omics data to resolve this problem.MethodsFive-omics data from 417 PCa patients from The Cancer Genome Atlas (TCGA) were used to construct the DL-based, relapse-sensitive model. Among them, 265 (63.5%) individuals experienced BCR. Five additional independent validation sets were applied to assess its predictive robustness. Bioinformatics analyses of two relapse-associated subgroups were then performed for identification of differentially expressed genes (DEGs), enriched pathway analysis, copy number analysis and immune cell infiltration analysis.ResultsThe DL-based model, with a significant difference (P = 6e-9) between two subgroups and good concordance index (C-index = 0.767), were proven to be robust by external validation. 1530 DEGs including 678 up- and 852 down-regulated genes were identified in the high-risk subgroup S2 compared with the low-risk subgroup S1. Enrichment analyses found five hallmark gene sets were up-regulated while 13 were down-regulated. Then, we found that DNA damage repair pathways were significantly enriched in the S2 subgroup. CNV analysis showed that 30.18% of genes were significantly up-regulated and gene amplification on chromosomes 7 and 8 was significantly elevated in the S2 subgroup. Moreover, enrichment analysis revealed that some DEGs and pathways were associated with immunity. Three tumor-infiltrating immune cell (TIIC) groups with a higher proportion in the S2 subgroup (p = 1e-05, p = 8.7e-06, p = 0.00014) and one TIIC group with a higher proportion in the S1 subgroup (P = 1.3e-06) were identified.ConclusionWe developed a novel, robust classification for understanding PCa relapse. This study validated the effectiveness of deep learning technique in prognosis prediction, and the method may benefit patients and prevent relapse by improving early detection and advancing early intervention.

An artificial intelligence model to predict 12-month mortality among patients with myocardial infarction

Abstract Funding Acknowledgements Type of funding sources: None. Background Ischemic heart disease and acute myocardial infarction (AMI) are the number one causes of death and disability, affecting around 126 million individuals worldwide [1]. About 30% of all cases of AMI lead to death. According to different sources, 8-20% of people with AMI die within 1 year of the event, and every second patient is rehospitalized [1-2]. Thus, it is necessary to assess the risk and to predict the outcomes during the first admission to the hospital, in order to guide the management and treatment after AMI [2]. Conventional risk rating scales and methods have a lot of drawbacks, so novel approaches such as artificial intelligence (AI) and deep learning (DL) methods have the potential to provide an accurate risk assessment and share decision-making in patients who experienced AMI [3-4]. Purpose To develop and validate a deep-learning-based risk stratification model to predict the 12-months mortality in patients with AMI. Methods A cohort of 250 AMI patients with Killip class I-IV was enrolled in the study. The data was collected on the admission day and a year after the event and included anamnesis as well as the results of the physical and laboratory-instrumental examination (72 values). All patients were randomly divided into 2 groups: the derivation cohort, which consisted of 215 individuals, and the validation cohort consisted of 35 patients. Our model was created using the Python programming language in GoogleColab and it is a convolutional neural network (CNN) with 1 input layer, 4 hidden layers, and 1 output layer consisting of neurons with a sigmoid activation function (Fig. 1). We used MaxAbsScaler for data normalization and Adam for optimization. Also, we calculated the GRACE risk score and compared the performance of the models only for the validation cohort that was not used for the model development. Results We developed our DL-based model using the derivation data during the 30 cycles of work, receiving the maximum value of training accuracy of 96.7%. The mortality risk, calculated by our model for the validation group was compared with real 12-months mortality, and the CNN "lied" only 2 times (Table 1). The test of the validation cohort showed such results of our DL-based model: 94% - accuracy, 71% - sensitivity, 100% - specificity, 100% - positive, and 93% - negative prognostic value. These results significantly outperformed the GRACE scale, which showed 63%, 71%, 54%, 26%, and 94% respectively. Conclusions In summary, we develop and validate a DL-based risk stratification model to predict the 12-months mortality in patients with AMI. The accuracy of the created model is significantly higher than conventional risk scales (GRACE score). Thus, it is proved that algorithms based on deep learning may be more effective to predict mortality and guide the management of cardiac patients.

Fully Automatic Quantitative Measurement of 18F-FDG PET/CT in Thymic Epithelial Tumors Using a Convolutional Neural Network.

The aim of this study was to develop a deep learning (DL)-based segmentation algorithm for automatic measurement of metabolic parameters of 18F-FDG PET/CT in thymic epithelial tumors (TETs), comparable performance to manual volumes of interest. A total of 186 consecutive patients with resectable TETs and preoperative 18F-FDG PET/CT were retrospectively enrolled (145 thymomas, 41 thymic carcinomas). A quasi-3D U-net architecture was trained to resemble ground-truth volumes of interest. Segmentation performance was assessed using the Dice similarity coefficient. Agreements between manual and DL-based automated extraction of SUVmax, metabolic tumor volume (MTV), total lesion glycolysis (TLG), and 63 radiomics features were evaluated via concordance correlation coefficients (CCCs) and linear regression slopes. Diagnostic and prognostic values were compared in terms of area under the receiver operating characteristics curve (AUC) for thymic carcinoma and hazards ratios (HRs) for freedom from recurrence. The mean Dice similarity coefficient was 0.83 ± 0.34. Automatically measured SUVmax (slope, 0.97; CCC, 0.92), MTV (slope, 0.94; CCC, 0.96), and TLG (slope, 0.96; CCC, 0.96) were in good agreement with manual measurements. The mean CCC and slopes were 0.88 ± 0.06 and 0.89 ± 0.05, respectively, for the radiomics parameters. Automatically measured SUVmax, MTV, and TLG showed good diagnostic accuracy for thymic carcinoma (AUCs: SUVmax, 0.95; MTV, 0.85; TLG, 0.87) and significant prognostic value (HRs: SUVmax, 1.31 [95% confidence interval, 1.16-1.48]; MTV, 2.11 [1.09-4.06]; TLG, 1.90 [1.12-3.23]). No significant differences in the AUCs or HRs were found between automatic and manual measurements for any of the metabolic parameters. Our DL-based model provides comparable segmentation performance and metabolic parameter values to manual measurements in TETs.

Improved diagnostic performance of plain radiography for cervical ossification of the posterior longitudinal ligament using deep learning.

BackgroundA high false-negative rate has been reported for the diagnosis of ossification of the posterior longitudinal ligament (OPLL) using plain radiography. We investigated whether deep learning (DL) can improve the diagnostic performance of radiologists for cervical OPLL using plain radiographs.Materials and methodsThe training set consisted of 915 radiographs from 207 patients diagnosed with OPLL. For the test set, we used 200 lateral cervical radiographs from 100 patients with cervical OPLL and 100 patients without OPLL. An observer performance study was conducted over two reading sessions. In the first session, we compared the diagnostic performance of the DL-model and the six observers. The diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC) at the vertebra and patient level. The sensitivity and specificity of the DL model and average observers were calculated in per-patient analysis. Subgroup analysis was performed according to the morphologic classification of OPLL. In the second session, observers evaluated the radiographs by referring to the results of the DL-model.ResultsIn the vertebra-level analysis, the DL-model showed an AUC of 0.854, which was higher than the average AUC of observers (0.826), but the difference was not significant (p = 0.292). In the patient-level analysis, the performance of the DL-model had an AUC of 0.851, and the average AUC of observers was 0.841 (p = 0.739). The patient-level sensitivity and specificity were 91% and 69% in the DL model, and 83% and 68% for the average observers, respectively. Both the DL-model and observers showed decreases in overall performance in the segmental and circumscribed types. With knowledge of the results of the DL-model, the average AUC of observers increased to 0.893 (p = 0.001) at the vertebra level and 0.911 (p < 0.001) at the patient level. In the subgroup analysis, the improvement was largest in segmental-type (AUC difference 0.087; p = 0.002).ConclusionsThe DL-based OPLL detection model can significantly improve the diagnostic performance of radiologists on cervical radiographs.

DeepBBBP: High Accuracy Blood-brain-barrier Permeability Prediction with a Mixed Deep Learning Model.

Blood-brain-barrier permeability (BBBP) is an important property that is used to establish the drug-likeness of a molecule, as it establishes whether the molecule can cross the BBB when desired. It also eliminates those molecules which are not supposed to cross the barrier, as doing so would lead to toxicity. BBBP can be measured in vivo, in vitro or in silico. With the advent and subsequent rise of in silico methods for virtual drug screening, quite a bit of work has been done to predict this feature using statistical machine learning (ML) and deep learning (DL) based methods. In this work a mixed DL-based model, consisting of a Multi-layer Perceptron (MLP) and Convolutional Neural Network layers, has been paired with Mol2vec. Mol2vec is a convenient and unsupervised machine learning technique which produces high-dimensional vector representations of molecules and its molecular substructures. These succinct vector representations are utilized as inputs to the mixed DL model that is used for BBBP predictions. Several well-known benchmarks incorporating BBBP data have been used for supervised training and prediction by our mixed DL model which demonstrates superior results when compared to existing ML and DL techniques used for predicting BBBP.

Evaluation Exploration of Atlas-Based and Deep Learning-Based Automatic Contouring for Nasopharyngeal Carcinoma.

PurposeThe purpose of this study was to evaluate and explore the difference between an atlas-based and deep learning (DL)-based auto-segmentation scheme for organs at risk (OARs) of nasopharyngeal carcinoma cases to provide valuable help for clinical practice.Methods120 nasopharyngeal carcinoma cases were established in the MIM Maestro (atlas) database and trained by a DL-based model (AccuContour®), and another 20 nasopharyngeal carcinoma cases were randomly selected outside the atlas database. The experienced physicians contoured 14 OARs from 20 patients based on the published consensus guidelines, and these were defined as the reference volumes (Vref). Meanwhile, these OARs were auto-contoured using an atlas-based model, a pre-built DL-based model, and an on-site trained DL-based model. These volumes were named Vatlas, VDL-pre-built, and VDL-trained, respectively. The similarities between Vatlas, VDL-pre-built, VDL-trained, and Vref were assessed using the Dice similarity coefficient (DSC), Jaccard coefficient (JAC), maximum Hausdorff distance (HDmax), and deviation of centroid (DC) methods. A one-way ANOVA test was carried out to show the differences (between each two of them).ResultsThe results of the three methods were almost similar for the brainstem and eyes. For inner ears and temporomandibular joints, the results of the pre-built DL-based model are the worst, as well as the results of atlas-based auto-segmentation for the lens. For the segmentation of optic nerves, the trained DL-based model shows the best performance (p < 0.05). For the contouring of the oral cavity, the DSC value of VDL-pre-built is the smallest, and VDL-trained is the most significant (p < 0.05). For the parotid glands, the DSC of Vatlas is the minimum (about 0.80 or so), and VDL-pre-built and VDL-trained are slightly larger (about 0.82 or so). In addition to the oral cavity, parotid glands, and the brainstem, the maximum Hausdorff distances of the other organs are below 0.5 cm using the trained DL-based segmentation model. The trained DL-based segmentation method behaves well in the contouring of all the organs that the maximum average deviation of the centroid is no more than 0.3 cm.ConclusionThe trained DL-based segmentation performs significantly better than atlas-based segmentation for nasopharyngeal carcinoma, especially for the OARs with small volumes. Although some delineation results still need further modification, auto-segmentation methods improve the work efficiency and provide a level of help for clinical work.

Development and validation of a deep learning model for detection of breast cancers in mammography from multi-institutional datasets.

ObjectivesThe objective of this study was to develop and validate a state-of-the-art, deep learning (DL)-based model for detecting breast cancers on mammography.MethodsMammograms in a hospital development dataset, a hospital test dataset, and a clinic test dataset were retrospectively collected from January 2006 through December 2017 in Osaka City University Hospital and Medcity21 Clinic. The hospital development dataset and a publicly available digital database for screening mammography (DDSM) dataset were used to train and to validate the RetinaNet, one type of DL-based model, with five-fold cross-validation. The model’s sensitivity and mean false positive indications per image (mFPI) and partial area under the curve (AUC) with 1.0 mFPI for both test datasets were externally assessed with the test datasets.ResultsThe hospital development dataset, hospital test dataset, clinic test dataset, and DDSM development dataset included a total of 3179 images (1448 malignant images), 491 images (225 malignant images), 2821 images (37 malignant images), and 1457 malignant images, respectively. The proposed model detected all cancers with a 0.45–0.47 mFPI and had partial AUCs of 0.93 in both test datasets.ConclusionsThe DL-based model developed for this study was able to detect all breast cancers with a very low mFPI. Our DL-based model achieved the highest performance to date, which might lead to improved diagnosis for breast cancer.

A new nearest neighbor-based framework for diabetes detection

Diabetes is one of the deadliest and costliest diseases. Today, automatic diabetes detection systems are primarily developed using deep learning (DL) approaches, which give high accuracy in classifying patients into two classes: have diabetes or not. Unfortunately, DL is a high-complexity and unexplainable black-box model. This paper proposes a new nearest neighbor-based framework to tackle those issues in classifying two diabetes datasets: binary-class Pima India Diabetes Dataset (PIDD) and multiclass Diabetes Type. A k-means clustering (KMC) is first carried out to remove the noises or outliers and keep the competent data in the training set. The dimension of the competent data is then reduced using an autoencoder (AE) to minimize the distances of the intra-class data but maximize that of the inter-class. A k-nearest neighbor (KNN) classifier and two variants: pseudo nearest neighbor rule (PNNR) and local mean-based pseudo nearest neighbor (LMPNN), are used to detect diabetes. In addition, a new variant named multi-voter multi-commission nearest neighbor (MVMCNN) is introduced. An investigation based on 5-fold cross-validation (FCV) informs that, for binary-class PIDD, the proposed combination of KMC, AE, and MVMCNN achieves the highest accuracy of 99.13%, which is slightly higher than the state-of-the-art DL-based detection model that produces 98.07%. An evaluation based on 10-FCV also indicates that, for the multiclass Diabetes Type, it obtains a higher accuracy of 95.24% than the DL-based model for predicting diabetes that gives 94.02%.

Automatic contouring of normal tissues with deep learning for preclinical radiation studies

Objective. Delineation of relevant normal tissues is a bottleneck in image-guided precision radiotherapy workflows for small animals. A deep learning (DL) model for automatic contouring using standardized 3D micro cone-beam CT (μCBCT) volumes as input is proposed, to provide a fully automatic, generalizable method for normal tissue contouring in preclinical studies. Approach. A 3D U-net was trained to contour organs in the head (whole brain, left/right brain hemisphere, left/right eye) and thorax (complete lungs, left/right lung, heart, spinal cord, thorax bone) regions. As an important preprocessing step, Hounsfield units (HUs) were converted to mass density (MD) values, to remove the energy dependency of the μCBCT scanner and improve generalizability of the DL model. Model performance was evaluated quantitatively by Dice similarity coefficient (DSC), mean surface distance (MSD), 95th percentile Hausdorff distance (HD95p), and center of mass displacement (ΔCoM). For qualitative assessment, DL-generated contours (for 40 and 80 kV images) were scored (0: unacceptable, manual re-contouring needed - 5: no adjustments needed). An uncertainty analysis using Monte Carlo dropout uncertainty was performed for delineation of the heart. Main results. The proposed DL model and accompanying preprocessing method provide high quality contours, with in general median DSC > 0.85, MSD < 0.25 mm, HD95p < 1 mm and ΔCoM < 0.5 mm. The qualitative assessment showed very few contours needed manual adaptations (40 kV: 20/155 contours, 80 kV: 3/155 contours). The uncertainty of the DL model is small (within 2%). Significance. A DL-based model dedicated to preclinical studies has been developed for multi-organ segmentation in two body sites. For the first time, a method independent of image acquisition parameters has been quantitatively evaluated, resulting in sub-millimeter performance, while qualitative assessment demonstrated the high quality of the DL-generated contours. The uncertainty analysis additionally showed that inherent model variability is low.

A Survey on Deep Learning for Software Engineering

In 2006, Geoffrey Hinton proposed the concept of training “Deep Neural Networks (DNNs)” and an improved model training method to break the bottleneck of neural network development. More recently, the introduction of AlphaGo in 2016 demonstrated the powerful learning ability of deep learning and its enormous potential. Deep learning has been increasingly used to develop state-of-the-art software engineering (SE) research tools due to its ability to boost performance for various SE tasks. There are many factors, e.g., deep learning model selection, internal structure differences, and model optimization techniques, that may have an impact on the performance of DNNs applied in SE. Few works to date focus on summarizing, classifying, and analyzing the application of deep learning techniques in SE. To fill this gap, we performed a survey to analyze the relevant studies published since 2006. We first provide an example to illustrate how deep learning techniques are used in SE. We then conduct a background analysis (BA) of primary studies and present four research questions to describe the trend of DNNs used in SE (BA), summarize and classify different deep learning techniques (RQ1), and analyze the data processing including data collection, data classification, data pre-processing, and data representation (RQ2). In RQ3, we depicted a range of key research topics using DNNs and investigated the relationships between DL-based model adoption and multiple factors (i.e., DL architectures, task types, problem types, and data types). We also summarized commonly used datasets for different SE tasks. In RQ4, we summarized the widely used optimization algorithms and provided important evaluation metrics for different problem types, including regression, classification, recommendation, and generation. Based on our findings, we present a set of current challenges remaining to be investigated and outline a proposed research road map highlighting key opportunities for future work.

Deep learning-based algorithm for lung cancer detection on chest radiographs using the segmentation method

We developed and validated a deep learning (DL)-based model using the segmentation method and assessed its ability to detect lung cancer on chest radiographs. Chest radiographs for use as a training dataset and a test dataset were collected separately from January 2006 to June 2018 at our hospital. The training dataset was used to train and validate the DL-based model with five-fold cross-validation. The model sensitivity and mean false positive indications per image (mFPI) were assessed with the independent test dataset. The training dataset included 629 radiographs with 652 nodules/masses and the test dataset included 151 radiographs with 159 nodules/masses. The DL-based model had a sensitivity of 0.73 with 0.13 mFPI in the test dataset. Sensitivity was lower in lung cancers that overlapped with blind spots such as pulmonary apices, pulmonary hila, chest wall, heart, and sub-diaphragmatic space (0.50–0.64) compared with those in non-overlapped locations (0.87). The dice coefficient for the 159 malignant lesions was on average 0.52. The DL-based model was able to detect lung cancers on chest radiographs, with low mFPI.

Enhanced Deep Learning Model for Personalized Cancer Treatment

Personalized medicine provides more safe and effective treatment by individualizing the choice of drug and dose based on an individual’s genetic profile. Cancer patients’ response to anti-cancer treatments (drugs) is one of the foremost challenges in personalized medicine that releases the target treatment. Both size and availability of drug sensitivity data have motivated researchers to develop Artificial Intelligence (AI), based models, for predicting drug response to advance cancer treatment. The concerned AI models include Machine Learning (ML) and the recently advanced Deep Learning (DL) based models. This paper introduces both; a data federation method and a DL-based model for predicting drug response. The fundamental goal is to generalize the predictor so it will be able to predict the response to different drugs accurately. As the data has a considerable effect on any AI model, the data federation is utilized to consolidate the data. The proposed consolidation process is carried out to make each cell line contains gene expression data, its mutation profile, and drug response data. ML models such as Support Vector Machine (SVM) and Linear Regression (LR) are used along with Principal Component Analysis (PCA) for feature reduction, and the AI models have been tested with and without data federation. The results show that data federation enhanced the accuracy and decreased the Mean Square Error (MSE) by almost 25%. The proposed DL model uses dimension reduction encoders. The encoder is a DL model that uses unsupervised learning. It is trained by integrating an encoder with a decoder to achieve equality between the input and output. The proposed model has achieved the best accuracy compared to some other recent models in terms of the Pearson correlation coefficient (PCC) as a performance measure. In addition, the results show that the Enhanced Deep Drug Response prediction (Enhanced Deep-DR) model has achieved the best PCC value even with the largest number of genes and drugs, which proves the high capacity and efficiency of the proposed model. Convolutional Neural Network (CNN) based-model is also implemented; it achieves higher accuracy in predicting the drug response than in some other DL-based models but less than the Enhanced Deep learning. The Enhanced Deep-DR achieves better accuracy within the range of 5% to 12% than other DL-models.

Automatic Music Classification Model Based on Instantaneous Frequency and CNNs in High Noise Environment.

Automatic music classification has significant research implications because it is the foundation for quick and efficient music resource retrieval and has a wide range of possible applications. In this study, DL is used to extract and categorize musical features, and a DL-based model for music feature extraction and classification is created. In this study, the instantaneous frequency and short-time Fourier transform are used to estimate the sine of a mixed music signal. Based on peak-frequency pairs, the DL algorithm is then used to calculate multiple candidate pitch estimates for each frame, and the melody pitch sequence is then obtained in accordance with the pitch profile duration and continuity characteristics. With this approach, the pitch can be calculated without reference to the fundamental frequency component. A music feature classification approach using spectrogram as input data and CNN as classifier is proposed at the same time in light of CNN's benefits in image processing. Studies reveal that this model's categorization and music feature extraction accuracy is as high as 94.18 percent and 95.66 percent, respectively. The outcomes demonstrate the efficiency of this technique for the extraction and classification of musical features. The field of music information retrieval is a good fit for it.

Prior-Guided Data Augmentation for Infrared Small Target Detection

Recently, a lot of deep learning (DL) methods have been proposed for infrared small target detection (ISTD). A DL-based model for the ISTD task needs large amounts of samples. However, the diversity of existing ISTD datasets is not sufficient to train a DL model with good generalization. To solve this issue, a data augmentation method called Prior-Guided Data Augmentation (PGDA) is proposed to expand the diversity of training samples indirectly without additional training data. Specifically, it decouples the target description and localization abilities by preserving the scale distribution and physical characteristics of targets. Furthermore, a multi-scene infrared small target dataset (MSISTD) consisting of 1077 images with 1343 instances is constructed. The number of images and the number of instances in MSISTD are 2.4 times and 2.5 times than those of the existing largest real ISTD dataset Single-frame Infrared Small Target (SIRST) Benchmark, respectively. Extensive experiments on the SIRST dataset and the constructed MSISTD dataset illustrate that the proposed PGDA improves the performance of existing DL-based ISTD methods without extra model complexity burdens. In comparison with SIRST, MSISTD has been evaluated as a more comprehensive and accurate benchmark for ISTD tasks.

VAEPass: A lightweight passwords guessing model based on variational auto-encoder

Password guessing has attracted considerable attention in recent years. With the successful application of deep learning (DL) methods in natural language processing, password guessing models that leverage deep learning techniques by treating passwords as short texts, e.g., Recurrent Neural Network-based model and PassGAN model, have been confirmed to be more effective in terms of generalizability. However, the architectures of these existing DL-based password guessing models are typically extremely complex, which makes the process of training and password generation time-consuming. It is desirable to build a lightweight password guessing model that can reduce the time required for model training while maintaining the password guessing effect.In this study, we propose VAEPass, a lightweight password guessing model based on a Variational Auto-Encoder (VAE), which comprises of an encoder and a decoder established using Gated Convolutional Neural Network (GCNN). Furthermore, we improve the proposed VAEPass model to treat common character combinations summarized from training passwords as tokens and guess passwords, known as VAEPasstoken, at a token-level. Experiments demonstrate that the matching rate of the proposed VAEPasstoken is 2.7%∼9.3% higher than that of PassGAN method in the one-site test. Moreover, compared with the state-of-the-art PassGAN model (i.e., a DL-based model), the parameters in VAEPass are approximately 32% of that in PassGAN and the training time required by VAEPass is approximately 11% of the time required by PassGAN.

Training, Validation, and Test of Deep Learning Models for Classification of Receptor Expressions in Breast Cancers From Mammograms.

The molecular subtype of breast cancer is an important component of establishing the appropriate treatment strategy. In clinical practice, molecular subtypes are determined by receptor expressions. In this study, we developed a model using deep learning to determine receptor expressions from mammograms. A developing data set and a test data set were generated from mammograms from the affected side of patients who were pathologically diagnosed with breast cancer from January 2006 through December 2016 and from January 2017 through December 2017, respectively. The developing data sets were used to train and validate the DL-based model with five-fold cross-validation for classifying expression of estrogen receptor (ER), progesterone receptor (PgR), and human epidermal growth factor receptor 2-neu (HER2). The area under the curves (AUCs) for each receptor were evaluated with the independent test data set. The developing data set and the test data set included 1,448 images (997 ER-positive and 386 ER-negative, 641 PgR-positive and 695 PgR-negative, and 220 HER2-enriched and 1,109 non-HER2-enriched) and 225 images (176 ER-positive and 40 ER-negative, 101 PgR-positive and 117 PgR-negative, and 53 HER2-enriched and 165 non-HER2-enriched), respectively. The AUC of ER-positive or -negative in the test data set was 0.67 (0.58-0.76), the AUC of PgR-positive or -negative was 0.61 (0.53-0.68), and the AUC of HER2-enriched or non-HER2-enriched was 0.75 (0.68-0.82). The DL-based model effectively classified the receptor expressions from the mammograms. Applying the DL-based model to predict breast cancer classification with a noninvasive approach would have additive value to patients.

Optimal ECG-lead selection increases generalizability of deep learning on ECG abnormality classification.

Deep learning (DL) has achieved promising performance in detecting common abnormalities from the 12-lead electrocardiogram (ECG). However, diagnostic redundancy exists in the 12-lead ECG, which could impose a systematic overfitting on DL, causing poor generalization. We, therefore, hypothesized that finding an optimal lead subset of the 12-lead ECG to eliminate the redundancy would help improve the generalizability of DL-based models. In this study, we developed and evaluated a DL-based model that has a feature extraction stage, an ECG-lead subset selection stage and a decision-making stage to automatically interpret multiple common ECG abnormality types. The data analysed in this study consisted of 6877 12-lead ECG recordings from CPSC 2018 (labelled as normal rhythm or eight types of ECG abnormalities, split into training (approx. 80%), validation (approx. 10%) and test (approx. 10%) sets) and 3998 12-lead ECG recordings from PhysioNet/CinC 2020 (labelled as normal rhythm or four types of ECG abnormalities, used as external text set). The ECG-lead subset selection module was introduced within the proposed model to efficiently constrain model complexity. It detected an optimal 4-lead ECG subset consisting of leads II, aVR, V1 and V4. The proposed model using the optimal 4-lead subset significantly outperformed the model using the complete 12-lead ECG on the validation set and on the external test dataset. The results demonstrated that our proposed model successfully identified an optimal subset of 12-lead ECG; the resulting 4-lead ECG subset improves the generalizability of the DL model in ECG abnormality interpretation. This study provides an outlook on what channels are necessary to keep and which ones may be ignored when considering an automated detection system for cardiac ECG abnormalities. This article is part of the theme issue 'Advanced computation in cardiovascular physiology: new challenges and opportunities'.