Real-world clinical scenarios during introduction of trastuzumab biosimilar for HER2-positive breast cancer in the European Union.

Aim: Trastuzumab-anns is an intravenously administered biosimilar to trastuzumab approved by the EMA and US FDA for treatment of HER2+ early and metastatic breast cancer as well as metastatic gastric cancer. Lack of real-world characterization of biosimilar use has hindered uptake. Methods: This observational chart review characterizes 488 patients who received trastuzumab-anns in EU clinical practice settings. Results: Approximately 2/3rds of patients initiated trastuzumab-anns in adjuvant and neoadjuvant settings and most were naive new starters (70%). 30% were switchers from another trastuzumab, among whom 48% switched from trastuzumab iv. reference product. Common reasons for trastuzumab-anns discontinuation were a switch to another biosimilar product (34.8%, n = 85) or to trastuzumab reference product (15.6%, n = 38). Conclusion: Trastuzumab-anns was widely used in various treatment settings for HER2+ breast cancer.

Some patients have a type of breast cancer caused by abnormal amounts of a normal growth factor receptor. This growth factor receptor, known as human epidermal growth factor receptor-2 (HER-2), plays a role in normal life changes that occur in breast tissue, including during pregnancy. HER-2 exists on the surface of breast cells and sends a signal inside cells for growth and proliferation. Sometimes an abnormal amount of HER-2 appears on breast cell surfaces, which causes HER-2 to promote excessive growth and proliferation and leads to HER2+ breast cancer. HER2+ breast cancer can be treated with trastuzumab, a medicine that specifically blocks HER-2 signals, and stops cancer cell growth. Trastuzumab has greatly improved outcomes for women worldwide with HER2+ breast cancer but trastuzumab is not always available due, in part, to its high cost. Biosimilars are medicines that are highly similar, but not identical, to the brand name (original) product and have been shown in clinical trials to result in no meaningful difference in efficacy and safety compared with the original product. Trastuzumab-anns is an intravenously administered biosimilar to trastuzumab. Biosimilars are as effective and safe as original products, although more cost-effective, such that physicians and patients can benefit from more information about their use in the real world. This study provided information about trastuzumab-anns use from clinical oncology practices in seven European countries. The study provides real world evidence that trastuzumab-anns is used widely across different patients with HER2+ breast cancer, including those with metastatic disease.

Design and Evaluation of NSAID Derivatives as AKR1C3 Inhibitors for Breast Cancer Treatment through Computer-Aided Drug Design and In Vitro Analysis.

Breast cancer is a major global health issue, causing high incidence and mortality rates as well as psychological stress for patients. Chemotherapy resistance is a common challenge, and the Aldo-keto reductase family one-member C3 enzyme is associated with resistance to anthracyclines like doxorubicin. Recent studies have identified celecoxib as a potential treatment for breast cancer. Virtual screening was conducted using a quantitative structure-activity relationship model to develop similar drugs; this involved backpropagation of artificial neural networks and structure-based virtual screening. The screening revealed that the C-6 molecule had a higher affinity for the enzyme (-11.4 kcal/mol), a lower half-maximal inhibitory concentration value (1.7 µM), and a safer toxicological profile than celecoxib. The compound C-6 was synthesized with an 82% yield, and its biological activity was evaluated. The results showed that C-6 had a more substantial cytotoxic effect on MCF-7 cells (62%) compared to DOX (63%) and celecoxib (79.5%). Additionally, C-6 had a less harmful impact on healthy L929 cells than DOX and celecoxib. These findings suggest that C-6 has promising potential as a breast cancer treatment.

Classifying Alzheimer's disease and normal subjects using machine learning techniques and genetic-environmental features.

BACKGROUND: Alzheimer's disease (AD) is complicated by multiple environmental and polygenetic factors. The accuracy of artificial neural networks (ANNs) incorporating the common factors for identifying AD has not been evaluated. METHODS: A total of 184 probable AD patients and 3773 healthy individuals aged 65 and over were enrolled. AD-related genes (51 SNPs) and 8 environmental factors were selected as features for multilayer ANN modeling. Random Forest (RF) and Support Vector Machine with RBF kernel (SVM) were also employed for comparison. Model results were verified using traditional statistics. RESULTS: The ANN achieved high accuracy (0.98), sensitivity (0.95), and specificity (0.96) in the intrinsic test for AD classification. Excluding age and genetic data still yielded favorable results (accuracy: 0.97, sensitivity: 0.94, specificity: 0.96). The assigned weights to ANN features highlighted the importance of mental evaluation, years of education, and specific genetic variations (CASS4 rs7274581, PICALM rs3851179, and TOMM40 rs2075650) for AD classification. Receiver operating characteristic analysis revealed AUC values of 0.99 (intrinsic test), 0.60 (TWB-GWA), and 0.72 (CG-WGS), with slightly lower AUC values (0.96, 0.80, 0.52) when excluding age in ANN. The performance of the ANN model in AD classification was comparable to RF, SVM (linear kernel), and SVM (RBF kernel). CONCLUSIONS: The ANN model demonstrated good sensitivity, specificity, and accuracy in AD classification. The top-weighted SNPs for AD prediction were CASS4 rs7274581, PICALM rs3851179, and TOMM40 rs2075650. The ANN model performed similarly to RF and SVM, indicating its capability to handle the complexity of AD as a disease entity.

Exploring a multi-output temporal convolutional network driven encoder-decoder framework for ammonia nitrogen forecasting.

Artificial neural networks exhibit significant advantages in terms of learning capability and generalizability, and have been increasingly applied in water quality prediction. Through learning a compressed representation of the input data, the Encoder-Decoder (ED) structure not only could remove noise and redundancies, but also could efficiently capture the complex nonlinear relationships of meteorological and water quality factors. The novelty of this study lies in proposing a multi-output Temporal Convolutional Network based ED model (TCN-ED) to make ammonia nitrogen forecasts for the first time. The contribution of our study is indebted to systematically assessing the significance of combining the ED structure with advanced neural networks for making accurate and reliable water quality forecasts. The water quality gauge station located at Haihong village of an island in Shanghai City of China constituted the case study. The model input contained one hourly water quality factor and hourly meteorological factors of 32 observed stations, where each factor was traced back to the previous 24 h and each meteorological factor of 32 gauge stations was aggregated into one areal average factor. A total of 13,128 hourly water quality and meteorological data were divided into two datasets corresponding to model training and testing stages. The Long Short-Term Memory based ED (LSTM-ED), LSTM and TCN models were constructed for comparison purposes. The results demonstrated that the developed TCN-ED model can succeed in mimicking the complex dependence between ammonia nitrogen and water quality and meteorological factors, and provide more accurate ammonia nitrogen forecasts (1- up to 6-h-ahead) than the LSTM-ED, LSTM and TCN models. The TCN-ED model, in general, achieved higher accuracy, stability and reliability compared with the other models. Consequently, the improvement can facilitate river water quality forecasting and early warning, as well as benefit water pollution prevention in the interest of river environmental restoration and sustainability.

Prediction of Daily Ambient Temperature and Its Hourly Estimation Using Artificial Neural Networks in an Agrometeorological Station in Castile and León, Spain.

This study evaluates the predictive modeling of the daily ambient temperature (maximum, Tmax; average, Tave; and minimum, Tmin) and its hourly estimation (T0h, …, T23h) using artificial neural networks (ANNs) for agricultural applications. The data, 2004-2010, were used for training and 2011 for validation, recorded at the SIAR agrometeorological station of Mansilla Mayor (León). ANN models for daily prediction have three neurons in the output layer (Tmax(t + 1), Tave(t + 1), Tmin(t + 1)). Two models were evaluated: (1) with three entries (Tmax(t), Tave(t), Tmin(t)), and (2) adding the day of the year (J(t)). The inclusion of J(t) improves the predictions, with an RMSE for Tmax = 2.56, Tave = 1.65 and Tmin = 2.09 (°C), achieving better results than the classical statistical methods (typical year Tave = 3.64 °C; weighted moving mean Tmax = 2.76, Tave = 1.81 and Tmin = 2.52 (°C); linear regression Tave = 1.85 °C; and Fourier Tmax = 3.75, Tave = 2.67 and Tmin = 3.34 (°C)) for one year. The ANN models for hourly estimation have 24 neurons in the output layer (T0h(t), …, T23h(t)) corresponding to the mean hourly temperature. In this case, the inclusion of the day of the year (J(t)) does not significantly improve the estimations, with an RMSE = 1.25 °C, but it improves the results of the ASHRAE method, which obtains an RMSE = 2.36 °C for one week. The results obtained, with lower prediction errors than those achieved with the classical methods, confirm the interest in using the ANN models for predicting temperatures in agricultural applications.

Daily Estimation of Global Solar Irradiation and Temperatures Using Artificial Neural Networks through the Virtual Weather Station Concept in Castilla and León, Spain.

In this article, the interpolation of daily data of global solar irradiation, and the maximum, average, and minimum temperatures were measured. These measurements were carried out in the agrometeorological stations belonging to the Agro-climatic Information System for Irrigation (SIAR, in Spanish) of the Region of Castilla and León, in Spain, through the concept of Virtual Weather Station (VWS), which is implemented with Artificial Neural Networks (ANNs). This is serving to estimate data in every point of the territory, according to their geographic coordinates (i.e., longitude and latitude). The ANNs of the Multilayer Feed-Forward Perceptron (MLP) used are daily trained, along with data recorded in 53 agro-meteorological stations, and where the validation of the results is conducted in the station of Tordesillas (Valladolid). The ANN models for daily interpolation were tested with one, two, three, and four neurons in the hidden layer, over a period of 15 days (from 1 to 15 June 2020), with a root mean square error (RMSE, MJ/m2) of 1.23, 1.38, 1.31, and 1.04, respectively, regarding the daily global solar irradiation. The interpolation of ambient temperature also performed well when applying the VWS concept, with an RMSE (°C) of 0.68 for the maximum temperature with an ANN of four hidden neurons, 0.58 for the average temperature with three hidden neurons, and 0.83 for the minimum temperature with four hidden neurons.

Neural Networks for Financial Time Series Forecasting.

Financial and economic time series forecasting has never been an easy task due to its sensibility to political, economic and social factors. For this reason, people who invest in financial markets and currency exchange are usually looking for robust models that can ensure them to maximize their profile and minimize their losses as much as possible. Fortunately, recently, various studies have speculated that a special type of Artificial Neural Networks (ANNs) called Recurrent Neural Networks (RNNs) could improve the predictive accuracy of the behavior of the financial data over time. This paper aims to forecast: (i) the closing price of eight stock market indexes; and (ii) the closing price of six currency exchange rates related to the USD, using the RNNs model and its variants: the Long Short-Term Memory (LSTM) and the Gated Recurrent Unit (GRU). The results show that the GRU gives the overall best results, especially for the univariate out-of-sample forecasting for the currency exchange rates and multivariate out-of-sample forecasting for the stock market indexes.

Hyperspectral Imaging for Bloodstain Identification.

Blood is key evidence to reconstruct crime scenes in forensic sciences. Blood identification can help to confirm a suspect, and for that reason, several chemical methods are used to reconstruct the crime scene however, these methods can affect subsequent DNA analysis. Therefore, this study presents a non-destructive method for bloodstain identification using Hyperspectral Imaging (HSI, 397-1000 nm range). The proposed method is based on the visualization of heme-components bands in the 500-700 nm spectral range. For experimental and validation purposes, a total of 225 blood (different donors) and non-blood (protein-based ketchup, rust acrylic paint, red acrylic paint, brown acrylic paint, red nail polish, rust nail polish, fake blood, and red ink) samples (HSI cubes, each cube is of size 1000 × 512 × 224, in which 1000 × 512 are the spatial dimensions and 224 spectral bands) were deposited on three substrates (white cotton fabric, white tile, and PVC wall sheet). The samples are imaged for up to three days to include aging. Savitzky Golay filtering has been used to highlight the subtle bands of all samples, particularly the aged ones. Based on the derivative spectrum, important spectral bands were selected to train five different classifiers (SVM, ANN, KNN, Random Forest, and Decision Tree). The comparative analysis reveals that the proposed method outperformed several state-of-the-art methods.

A Data Fusion Method for Non-Destructive Testing by Means of Artificial Neural Networks.

In the aeronautics sector, aircraft parts are inspected during manufacture, assembly and service, to detect defects eventually present. Defects can be of different types, sizes and orientations, appearing in materials presenting a complex structure. Among the different inspection techniques, Non Destructive Testing (NDT) presents several advantages as they are noninvasive and cost effective. Within the NDT methods, Ultrasonic (US) waves are widely used to detect and characterize defects. However, due the so-called blind zone, they cannot be easily employed for defects close to the surface being inspected. On the other hand, another NDT technique such Eddy Current (EC) can be used only for detecting flaws close to the surface, due to the presence of the EC skin effect. The work presented in this article aims to combine the use of these two NDT methods, exploiting their complementary advantages. To reach this goal, a data fusion method is developed, by using Machine Learning techniques such as Artificial Neural Networks (ANNs). A simulated training database involving simulations of US and EC signals propagating in an Aluminum block in the presence of Side Drill Holes (SDHs) has been implemented, to train the ANNs. Measurements have been then performed on an Aluminum block, presenting tree different SDHs at specific depths. The trained ANNs were used to characterize the different real SDHs, providing an experimental validation. Eventually, particular attention has been addressed to the estimation errors corresponding to each flaw. Experimental results will show that depths and radii estimations error were confined on average within a range of 4%, recording a peak of 11% for the second SDHs.

Artificial Neural Networks in Computer-Aided Drug Design: An Overview of Recent Advances.

Computer-aided drug design (CADD) is the framework in which the huge amount of data accumulated by high-throughput experimental methods used in drug design is quantitatively studied. Its objectives include pattern recognition, biomarker identification and/or classification, etc. In order to achieve these objectives, machine learning algorithms and especially artificial neural networks (ANNs) have been used over ADMET factor testing and QSAR modeling evaluation. This paper provides an overview of the current trends in CADD-applied ANNs, since their use was re-boosted over a decade ago.

Characterization and Differentiation between Olive Varieties through Electrical Impedance Spectroscopy, Neural Networks and IoT.

Electrical impedance has shown itself to be useful in measuring the properties and characteristics of agri-food products: fruit quality, moisture content, the germination capacity in seeds or the frost-resistance of fruit. In the case of olives, it has been used to determine fat content and optimal harvest time. In this paper, a system based on the System on Chip (SoC) AD5933 running a 1024-point discrete Fourier transform (DFT) to return the impedance value as a magnitude and phase and which, working together with two ADG706 analog multiplexers and an external programmable clock based on a synthesized DDS in a FPGA XC3S250E-4VQG100C, allows for the impedance measurement in agri-food products with a frequency sweep from 1 Hz to 100 kHz. This paper demonstrates how electrical impedance is affected by the temperature both in freshly picked olives and in those processed in brine and provides a way to characterize cultivars by making use of only the electrical impedance, neural networks (NN) and the Internet of Things (IoT), allowing information to be collected from the olive samples analyzed both on farms and in factories.

Analysis of the Functionality of the Feed Chain in Olive Pitting, Slicing and Stuffing Machines by IoT, Computer Vision and Neural Network Diagnosis.

Olive pitting, slicing and stuffing machines (DRR in Spanish) are characterized by the fact that their optimal functioning is based on appropriate adjustments. Traditional systems are not completely reliable because their minimum error rate is 1-2%, which can result in fruit loss, since the pitting process is not infallible, and food safety issues can arise. Such minimum errors are impossible to remove through mechanical adjustments. In order to achieve this objective, an innovative solution must be provided in order to remove errors at operating speed rates over 2500 olives/min. This work analyzes the appropriate placement of olives in the pockets of the feed chain by using the following items: (1) An IoT System to control the DRR machine and the data analysis. (2) A computer vision system with an external shot camera and a LED lighting system, which takes a picture of every pocket passing in front of the camera. (3) A chip with a neural network for classification that, once trained, classifies between four possible pocket cases: empty, normal, incorrectly de-stoned olives at any angles (also known as a "boat"), and an anomalous case (foreign elements such as leafs, small branches or stones, two olives or small parts of olives in the same pocket). The main objective of this paper is to illustrate how with the use of a system based on IoT and a physical chip (NeuroMem CM1K, General Vision Inc.) with neural networks for sorting purposes, it is possible to optimize the functionality of this type of machine by remotely analyzing the data obtained. The use of classifying hardware allows it to work at the nominal operating speed for these machines. This would be limited if other classifying techniques based on software were used.

The relationship between air pollution and COVID-19-related deaths: An application to three French cities.

Being heavily dependent to oil products (mainly gasoline and diesel), the French transport sector is the main emitter of Particulate Matter (PMs) whose critical levels induce harmful health effects for urban inhabitants. We selected three major French cities (Paris, Lyon, and Marseille) to investigate the relationship between the Coronavirus Disease 19 (COVID-19) outbreak and air pollution. Using Artificial Neural Networks (ANNs) experiments, we have determined the concentration of PM2.5 and PM10 linked to COVID-19-related deaths. Our focus is on the potential effects of Particulate Matter (PM) in spreading the epidemic. The underlying hypothesis is that a pre-determined particulate concentration can foster COVID-19 and make the respiratory system more susceptible to this infection. The empirical strategy used an innovative Machine Learning (ML) methodology. In particular, through the so-called cutting technique in ANNs, we found new threshold levels of PM2.5 and PM10 connected to COVID-19: 17.4 µg/m3 (PM2.5) and 29.6 µg/m3 (PM10) for Paris; 15.6 µg/m3 (PM2.5) and 20.6 µg/m3 (PM10) for Lyon; 14.3 µg/m3 (PM2.5) and 22.04 µg/m3 (PM10) for Marseille. Interestingly, all the threshold values identified by the ANNs are higher than the limits imposed by the European Parliament. Finally, a Causal Direction from Dependency (D2C) algorithm is applied to check the consistency of our findings.

Non-Linear Quantitative Structure⁻Activity Relationships Modelling, Mechanistic Study and In-Silico Design of Flavonoids as Potent Antioxidants.

In this work, we developed quantitative structure-activity relationships (QSAR) models for prediction of oxygen radical absorbance capacity (ORAC) of flavonoids. Both linear (partial least squares-PLS) and non-linear models (artificial neural networks-ANNs) were built using parameters of two well-established antioxidant activity mechanisms, namely, the hydrogen atom transfer (HAT) mechanism defined with the minimum bond dissociation enthalpy, and the sequential proton-loss electron transfer (SPLET) mechanism defined with proton affinity and electron transfer enthalpy. Due to pronounced solvent effects within the ORAC assay, the hydration energy was also considered. The four-parameter PLS-QSAR model yielded relatively high root mean square errors (RMSECV = 0.783, RMSEE = 0.668, RMSEP = 0.900). Conversely, the ANN-QSAR model yielded considerably lower errors (RMSEE = 0.180 ± 0.059, RMSEP1 = 0.164 ± 0.128, and RMSEP2 = 0.151 ± 0.114) due to the inherent non-linear relationships between molecular structures of flavonoids and ORAC values. Five-fold cross-validation was found to be unsuitable for the internal validation of the ANN-QSAR model with a high RMSECV of 0.999 ± 0.253; which is due to limited sample size where resampling with replacement is a considerably better alternative. Chemical domains of applicability were defined for both models confirming their reliability and robustness. Based on the PLS coefficients and partial derivatives, both models were interpreted in terms of the HAT and SPLET mechanisms. Theoretical computations based on density functional theory at ωb97XD/6-311++G(d,p) level of theory were also carried out to further shed light on the plausible mechanism of anti-peroxy radical activity. Calculated energetics for simplified models (genistein and quercetin) with peroxyl radical derived from 2,2'-azobis (2-amidino-propane) dihydrochloride suggested that both SPLET and single electron transfer followed by proton loss (SETPL) mechanisms are competitive and more favorable than HAT in aqueous medium. The finding is in good accord with the ANN-based QSAR modelling results. Finally, the strongly predictive ANN-QSAR model was used to predict antioxidant activities for a series of 115 flavonoids designed combinatorially with flavone as a template. Structural trends were analyzed, and general guidelines for synthesis of new flavonoid derivatives with potentially potent antioxidant activities were given.

Interpretation of ANN-based QSAR models for prediction of antioxidant activity of flavonoids.

Quantitative structure-activity relationships (QSARs) built using machine learning methods, such as artificial neural networks (ANNs) are powerful in prediction of (antioxidant) activity from quantum mechanical (QM) parameters describing the molecular structure, but are usually not interpretable. This obvious difficulty is one of the most common obstacles in application of ANN-based QSAR models for design of potent antioxidants or elucidating the underlying mechanism. Interpreting the resulting models is often omitted or performed erroneously altogether. In this work, a comprehensive comparative study of six methods (PaD, PaD2 , weights, stepwise, perturbation and profile) for exploration and interpretation of ANN models built for prediction of Trolox-equivalent antioxidant capacity (TEAC) QM descriptors, is presented. Sum of ranking differences (SRD) was used for ranking of the six methods with respect to the contributions of the calculated QM molecular descriptors toward TEAC. The results show that the PaD, PaD2 and profile methods are the most stable and give rise to realistic interpretation of the observed correlations. Therefore, they are safely applicable for future interpretations without the opinion of an experienced chemist or bio-analyst. © 2018 Wiley Periodicals, Inc.

Developing a Mixed Neural Network Approach to Forecast the Residential Electricity Consumption Based on Sensor Recorded Data.

In this paper, we report a study having as a main goal the obtaining of a method that can provide an accurate forecast of the residential electricity consumption, refining it up to the appliance level, using sensor recorded data, for residential smart homes complexes that use renewable energy sources as a part of their consumed electricity, overcoming the limitations of not having available historical meteorological data and the unwillingness of the contractor to acquire such data periodically in the future accurate short-term forecasts from a specialized institute due to the implied costs. In this purpose, we have developed a mixed artificial neural network (ANN) approach using both non-linear autoregressive with exogenous input (NARX) ANNs and function fitting neural networks (FITNETs). We have used a large dataset containing detailed electricity consumption data recorded by sensors, monitoring a series of individual appliances, while in the NARX case we have also used timestamps datasets as exogenous variables. After having developed and validated the forecasting method, we have compiled it in view of incorporating it into a cloud solution, being delivered to the contractor that can provide it as a service for a monthly fee to both the operators and residential consumers.

Identification of Plasma Waves at Saturn Using Convolutional Neural Networks.

Cassini has recently completed its 13-year mission at Saturn leaving a vast data set. A large interest among the scientific community is to investigate plasma waves and instabilities at Saturn. It is no longer feasible to manually search through Cassini's vast data set to identify all such waves of interest. Thus, the feasibility of using artificial neural networks (ANNs) to identify plasma waves at Saturn is demonstrated using Cassini data. A convolutional neural network (CNN) was trained to identify low-frequency plasma waves that occur in the upstream region of Saturn using images constructed from the Cassini magnetometer time series data. By systematically varying the network architecture during training and validation, a CNN was obtained that can identify upstream waves with an accuracy of 94% ± 2%. The CNN's high accuracy for wave identification demonstrates that it is, in fact, feasible to use ANNs to identify plasma waves at Saturn and by extension in other planetary and lunar plasma environments using spacecraft data.

Prediction of metastasis in advanced colorectal carcinomas using CGH data.

Logistic Regression Model (LRM) and artificial neural networks (ANNs) as two nonlinear models have been used to establish a novel two-stage hybrid modeling procedure for prediction of metastasis in advanced colorectal carcinomas. Two different datasets were used in training and testing procedures. For the first stage of hybrid modeling procedure, LRM was used to evaluate the contribution of DNA sequence copy number aberrations detected by Comparative Genomic Hybridization in advanced colorectal carcinoma and its metastasis. Then, the most effective parameters were selected by the LRM. Selected effective parameters among 565 detected chromosomal gains and losses were as follows: gain of 20q11.2, loss of 1q42, loss of 13q34, gain of 5q12, gain of 17p13, loss of 2q22, loss of 11q24 and gain of 2p11.2. Consequently, neural network models were constructed and fed by the parameters selected by LRM to build hybrid predictors on the two databases during self-consistency and jackknife tests, and performance of the hybrid model was verified. The results showed that our two-stage hybrid model approach is very promising for prediction of metastasis in advanced colorectal carcinomas.

Deep Convolutional Neural Networks for Endotracheal Tube Position and X-ray Image Classification: Challenges and Opportunities.

The goal of this study is to evaluate the efficacy of deep convolutional neural networks (DCNNs) in differentiating subtle, intermediate, and more obvious image differences in radiography. Three different datasets were created, which included presence/absence of the endotracheal (ET) tube (n = 300), low/normal position of the ET tube (n = 300), and chest/abdominal radiographs (n = 120). The datasets were split into training, validation, and test. Both untrained and pre-trained deep neural networks were employed, including AlexNet and GoogLeNet classifiers, using the Caffe framework. Data augmentation was performed for the presence/absence and low/normal ET tube datasets. Receiver operating characteristic (ROC), area under the curves (AUC), and 95% confidence intervals were calculated. Statistical differences of the AUCs were determined using a non-parametric approach. The pre-trained AlexNet and GoogLeNet classifiers had perfect accuracy (AUC 1.00) in differentiating chest vs. abdominal radiographs, using only 45 training cases. For more difficult datasets, including the presence/absence and low/normal position endotracheal tubes, more training cases, pre-trained networks, and data-augmentation approaches were helpful to increase accuracy. The best-performing network for classifying presence vs. absence of an ET tube was still very accurate with an AUC of 0.99. However, for the most difficult dataset, such as low vs. normal position of the endotracheal tube, DCNNs did not perform as well, but achieved a reasonable AUC of 0.81.

Detection and Labeling of Vertebrae in MR Images Using Deep Learning with Clinical Annotations as Training Data.

The purpose of this study was to investigate the potential of using clinically provided spine label annotations stored in a single institution image archive as training data for deep learning-based vertebral detection and labeling pipelines. Lumbar and cervical magnetic resonance imaging cases with annotated spine labels were identified and exported from an image archive. Two separate pipelines were configured and trained for lumbar and cervical cases respectively, using the same setup with convolutional neural networks for detection and parts-based graphical models to label the vertebrae. The detection sensitivity, precision and accuracy rates ranged between 99.1-99.8, 99.6-100, and 98.8-99.8% respectively, the average localization error ranges were 1.18-1.24 and 2.38-2.60 mm for cervical and lumbar cases respectively, and with a labeling accuracy of 96.0-97.0%. Failed labeling results typically involved failed S1 detections or missed vertebrae that were not fully visible on the image. These results show that clinically annotated image data from one image archive is sufficient to train a deep learning-based pipeline for accurate detection and labeling of MR images depicting the spine. Further, these results support using deep learning to assist radiologists in their work by providing highly accurate labels that only require rapid confirmation.