Unseen Test Dataset Research Articles

Comparative study on total nitrogen prediction in wastewater treatment plant and effect of various feature selection methods on machine learning algorithms performance

Wastewater characteristics prediction in wastewater treatment plants (WWTPs) is valuable and can reduce the number of sampling, energy, and cost. Feature Selection (FS) methods are used in the pre-processing section for enhancing the model performance. This study aims to evaluate the effect of seven different FS methods (filter, wrapper, and embedded methods) on enhancing the prediction accuracy for total nitrogen (TN) in the WWTP influent flow. Four scenarios based on FS suggestions were defined and compared by three supervised Machine Learning (ML) algorithms, i.e. Artificial Neural Network (ANN), Random Forest (RF), and Gradient Boosting Machine (GBM). Input parameters, as daily time-series including pH, DO, COD, BOD, MLSS, MLVSS, NH4-N, and TN concentration, were used. Data set divided into train and unseen test data-sets, and performance precision of all models was carried out based on Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and correlation coefficient (R2). Results reveal that scenario IV which was suggested by Mutual Information, including NH4-N, COD, BOD, and DO had the best result rather than other FS methods. Furthermore, decision tree algorithms (RF and GBM) revealed better performance results in comparison to neural network algorithm (ANN). GBM generalized the dataset patterns very well and produced the best performance on unseen data-set, which shows the effectiveness of this state-of-the-art ML algorithm for wastewater components prediction.

Grading hypoxic-ischemic encephalopathy in neonatal EEG with convolutional neural networks and quadratic time–frequency distributions

Objective. To develop an automated system to classify the severity of hypoxic-ischaemic encephalopathy injury (HIE) in neonates from the background electroencephalogram (EEG). Approach. By combining a quadratic time–frequency distribution (TFD) with a convolutional neural network, we develop a system that classifies 4 EEG grades of HIE. The network learns directly from the two-dimensional TFD through 3 independent layers with convolution in the time, frequency, and time–frequency directions. Computationally efficient algorithms make it feasible to transform each 5 min epoch to the time–frequency domain by controlling for oversampling to reduce both computation and computer memory. The system is developed on EEG recordings from 54 neonates. Then the system is validated on a large unseen dataset of 338 h of EEG recordings from 91 neonates obtained across multiple international centres. Main results. The proposed EEG HIE-grading system achieves a leave-one-subject-out testing accuracy of 88.9% and kappa of 0.84 on the development dataset. Accuracy for the large unseen test dataset is 69.5% (95% confidence interval, CI: 65.3%–73.6%) and kappa of 0.54, which is a significant () improvement over a state-of-the-art feature-based method with an accuracy of 56.8% (95% CI: 51.4%–61.7%) and kappa of 0.39. Performance of the proposed system was unaffected when the number of channels in testing was reduced from 8 to 2—accuracy for the large validation dataset remained at 69.5% (95% CI: 65.5%–74.0%). Significance. The proposed system outperforms the state-of-the-art machine learning algorithms for EEG grade classification on a large multi-centre unseen dataset, indicating the potential to assist clinical decision making for neonates with HIE.

Prediction of heavy metals removal by polymer inclusion membranes using machine learning techniques

AbstractThis study predicts heavy metals removal from aqueous solution by polymer inclusion membranes (PIMs) process using machine learning (ML) techniques such as multiple layer perceptron neural networks (MLPNN) and multiple linear regression (MLR) after data analysis. The removal efficiency (RE) of the PIMs process is predicted for cobalt (Co), cadmium (Cd) and chromium (Cr) by changing operating conditions including time, carrier type, carrier rate, film thickness, plasticizer type and plasticizer rate. The MLPNN model presents reliable results with lower mean square error (MSE) for an unseen testing dataset, whereas the MLR model shows higher MSE values. The coefficient of determination (R2) of the MLPNN model for the testing dataset is 0.93, 0.90 and 0.86 for Co, Cd and Cr, respectively, whereas MLR shows poor results. Therefore, the MLPNN model can be a competitive, robust and fast alternate to optimize the PIMs process with minimum experimental work.

Abstract PR-08: Accurate quantification of tumor DNA in liquid biopsies using deep learning

Abstract Circulating tumor DNA (ctDNA) fraction has been shown to be a prognostic factor of treatment response in patients with metastatic cancers. However, algorithms that accurately determine ctDNA fraction across purity ranges are still lacking. Commonly used liquid biopsy gene panels (~70 genes) often approximate the ctDNA fraction in a sample by using somatic variant allele frequencies, resulting in large confidence intervals due to the scarcity of somatic mutations detected per sample and uncertainty of their clonality and copy number. Genome wide signatures such as nucleosome positioning, measured by windowed protection scores (WPS), have been shown to reflect genomic and epigenomic abnormalities specific to cancer and proposed as biomarkers for cancer detection. However, genome wide assays sacrifice somatic mutation detection due to low sequencing coverage and in clinical practice a second assay is required to find actionable mutations. We leverage data from ~6,000 cell-free DNA samples to develop methods for accurately estimating the ctDNA fraction in clinical samples sequenced with a 554-gene panel covering 3MB of genomic content. First, we use Gradient Boosting and 17 features extracted from standard variant calling, copy number and QC pipelines. This approach has been shown to be highly accurate in samples of high ctDNA fraction and therefore provide reliable training labels for those samples. We suggest that more sophisticated methods, using genome wide signatures, can increase prediction accuracies in samples of lower ctDNA fraction. Second, we develop a deep learning model for predicting ctDNA fraction using base pair resolution WPS over the genomic territory covered by the gene panel. Using Morlet wavelet, we transformed each autosomal WPS unidimensional representation into a two-dimensional image. A 22 channels image per sample was generated and used as input to a Convolutional Neural Network (22-CNN) model. The 22-CNN was trained to predict the ctDNA fraction and to classify if a sample was from a ctDNA shedder (ctDNA fraction >0) or not. The dataset was equally divided into training and test sets of ~2800 samples each. We used five-fold cross-validation (CV) during training and evaluated each model in the unseen test dataset. The Mean Square Error of the 22-CNN ctDNA prediction model on the testing set was 0.0072 +- 0.0013 [0.006-0.009] and the accuracy of the classification model on the testing set was 0.79 +- 0.006 [0.78-0.79]. In this work, we present both machine learning and deep learning models to reliably estimate the ctDNA fraction in plasma samples from cancer patients. These methods address a key challenge in the field and are applicable to other liquid biopsy assays. Citation Format: Aamna M. Al-Shehhi, Chintan Parmar, Ben Gustafson, Angad Singh, Rebecca Leary, Markus Riester, O. Alejandro Balbin. Accurate quantification of tumor DNA in liquid biopsies using deep learning [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PR-08.

Artificial intelligence-based predictive model of nanoscale friction using experimental data

A recent systematic experimental characterisation of technological thin films, based on elaborated design of experiments as well as probe calibration and correction procedures, allowed for the first time the determination of nanoscale friction under the concurrent influence of several process parameters, comprising normal forces, sliding velocities, and temperature, thus providing an indication of the intricate correlations induced by their interactions and mutual effects. This created the preconditions to undertake in this work an effort to model friction in the nanometric domain with the goal of overcoming the limitations of currently available models in ascertaining the effects of the physicochemical processes and phenomena involved in nanoscale contacts. Due to the stochastic nature of nanoscale friction and the relatively sparse available experimental data, meta-modelling tools fail, however, at predicting the factual behaviour. Based on the acquired experimental data, data mining, incorporating various state-of-the-art machine learning (ML) numerical regression algorithms, is therefore used. The results of the numerical analyses are assessed on an unseen test dataset via a comparative statistical validation. It is therefore shown that the black box ML methods provide effective predictions of the studied correlations with rather good accuracy levels, but the intrinsic nature of such algorithms prevents their usage in most practical applications. Genetic programming-based artificial intelligence (AI) methods are consequently finally used. Despite the marked complexity of the analysed phenomena and the inherent dispersion of the measurements, the developed AI-based symbolic regression models allow attaining an excellent predictive performance with the respective prediction accuracy, depending on the sample type, between 72% and 91%, allowing also to attain an extremely simple functional description of the multidimensional dependence of nanoscale friction on the studied variable process parameters. An effective tool for nanoscale friction prediction, adaptive control purposes, and further scientific and technological nanotribological analyses is thus obtained.

COVID-DeepPredictor: Recurrent Neural Network to Predict SARS-CoV-2 and Other Pathogenic Viruses.

The COVID-19 disease for Novel coronavirus (SARS-CoV-2) has turned out to be a global pandemic. The high transmission rate of this pathogenic virus demands an early prediction and proper identification for the subsequent treatment. However, polymorphic nature of this virus allows it to adapt and sustain in different kinds of environment which makes it difficult to predict. On the other hand, there are other pathogens like SARS-CoV-1, MERS-CoV, Ebola, Dengue, and Influenza as well, so that a predictor is highly required to distinguish them with the use of their genomic information. To mitigate this problem, in this work COVID-DeepPredictor is proposed on the framework of deep learning to identify an unknown sequence of these pathogens. COVID-DeepPredictor uses Long Short Term Memory as Recurrent Neural Network for the underlying prediction with an alignment-free technique. In this regard, k-mer technique is applied to create Bag-of-Descriptors (BoDs) in order to generate Bag-of-Unique-Descriptors (BoUDs) as vocabulary and subsequently embedded representation is prepared for the given virus sequences. This predictor is not only validated for the dataset using -fold cross-validation but also for unseen test datasets of SARS-CoV-2 sequences and sequences from other viruses as well. To verify the efficacy of COVID-DeepPredictor, it has been compared with other state-of-the-art prediction techniques based on Linear Discriminant Analysis, Random Forests, and Gradient Boosting Method. COVID-DeepPredictor achieves 100% prediction accuracy on validation dataset while on test datasets, the accuracy ranges from 99.51 to 99.94%. It shows superior results over other prediction techniques as well. In addition to this, accuracy and runtime of COVID-DeepPredictor are considered simultaneously to determine the value of k in k-mer, a comparative study among k values in k-mer, Bag-of-Descriptors (BoDs), and Bag-of-Unique-Descriptors (BoUDs) and a comparison between COVID-DeepPredictor and Nucleotide BLAST have also been performed. The code, training, and test datasets used for COVID-DeepPredictor are available at http://www.nitttrkol.ac.in/indrajit/projects/COVID-DeepPredictor/.

Prediction of treatment response to transarterial radioembolization of liver metastases: Radiomics analysis of pre-treatment cone-beam CT: A proof of concept study

PurposeTo investigate the potential of texture analysis and machine learning to predict treatment response to transarterial radioembolization (TARE) on pre-interventional cone-beam computed tomography (CBCT) images in patients with liver metastases. Materials and MethodsIn this IRB-approved retrospective single-center study 36 patients with a total of 104 liver metastases (56 % male, mean age 61.1 ± 13 years) underwent CBCT prior to TARE and follow-up imaging 6 months after therapy. Treatment response was evaluated according to RECIST version 1.1 and dichotomized into disease control (partial response/stable disease) versus disease progression (progressive disease). After target lesion segmentation, 104 radiomics features corresponding to seven different feature classes were extracted with the pyRadiomics package. After dimension reduction machine learning classifications were performed on a custom artificial neural network (ANN). Ten-fold cross validation on a previously unseen test data set was performed. ResultsThe average administered cumulative activity from TARE was 1.6 Gbq (± 0.5 Gbq). At a mean follow-up of 5.9 ± 0.8 months disease control was achieved in 82 % of metastases. After dimension reduction, 15 of 104 (15 %) texture analysis features remained for further analysis. On a previously unseen set of liver metastases the Multilayer Perceptron ANN yielded a sensitivity of 94.2 %, specificity of 67.7 % and an area-under-the receiver operating characteristics curve of 0.85. ConclusionOur study indicates that texture analysis-based machine learning may has potential to predict treatment response to TARE using pre-treatment CBCT images of patients with liver metastases with high accuracy.

DoFE: Domain-Oriented Feature Embedding for Generalizable Fundus Image Segmentation on Unseen Datasets.

Deep convolutional neural networks have significantly boosted the performance of fundus image segmentation when test datasets have the same distribution as the training datasets. However, in clinical practice, medical images often exhibit variations in appearance for various reasons, e.g., different scanner vendors and image quality. These distribution discrepancies could lead the deep networks to over-fit on the training datasets and lack generalization ability on the unseen test datasets. To alleviate this issue, we present a novel Domain-oriented Feature Embedding (DoFE) framework to improve the generalization ability of CNNs on unseen target domains by exploring the knowledge from multiple source domains. Our DoFE framework dynamically enriches the image features with additional domain prior knowledge learned from multi-source domains to make the semantic features more discriminative. Specifically, we introduce a Domain Knowledge Pool to learn and memorize the prior information extracted from multi-source domains. Then the original image features are augmented with domain-oriented aggregated features, which are induced from the knowledge pool based on the similarity between the input image and multi-source domain images. We further design a novel domain code prediction branch to infer this similarity and employ an attention-guided mechanism to dynamically combine the aggregated features with the semantic features. We comprehensively evaluate our DoFE framework on two fundus image segmentation tasks, including the optic cup and disc segmentation and vessel segmentation. Our DoFE framework generates satisfying segmentation results on unseen datasets and surpasses other domain generalization and network regularization methods.

Smart Handheld Based Human Activity Recognition Using Multiple Instance Multiple Label Learning

Human activity recognition (HAR) and monitoring is beneficial for many medical applications, such as eldercare and post-trauma rehabilitation after surgery. HAR models based on smartphone’s accelerometer data could provide a convenient and ubiquitous solution to this problem. However, such models are mostly concerned with identifying basic activities such as ‘stand’/‘walk’ and thus the high-level context such as ‘walk in a queue’ for which a set of specific activities is performed remain unnoticed. Consequently, in this paper, we design a HAR framework that can identify a group of activities (rather than a single basic activity) being performed in a time window, thus, enables us to extract more meaningful information about the subject’s overall context. An algorithm is designed to formulate HAR as a multi-instance multi-label (MIML) learning problem. The procedure of generating feature bags of consecutive activity traces having multiple labels is formulated. In this work, the temporal relationship among activities is exploited to obtain a more comprehensive HAR model. Interestingly, the framework is found to completely/partially identify activity sequences that may not even be present in the training dataset. The framework is implemented and found to be working adequately when tested with real dataset collected from 8 users for 12 different activity combinations. MIML-kNN is found to provide maximum average precision (around 90%) even for an unseen test data-set.

Data-driven hybrid model and operating algorithm to shave peak demand costs of building electricity

Forecasting technologies aiming to reduce building electricity consumption and costs have recently been developed. However, to design a more accurate forecasting model using a physical-based approach, complicated mathematical modeling and multi-variable parameters and constraints must be considered. Moreover, single data-driven models usually have a high generalization error regarding previously unseen data. Most hybrid forecasting models usually do not consider shaving the peak demand costs based on time of use (TOU) to save building electricity costs. To overcome this limitation, we propose a data-driven hybrid forecasting model and an operating algorithm to shave the peak demand costs based on the TOU. The proposed model and algorithm are expected to reduce peak demand costs by 22%. In addition, the experimental results indicate that the hybrid model can reduce generalization errors because it improves the forecasting performance, achieving the highest recall (91.01%) among other models using a previously unseen test dataset.

Stroke prognostication for discharge planning with machine learning: A derivation study

Post-stroke discharge planning may be aided by accurate early prognostication. Machine learning may be able to assist with such prognostication. The study’s primary aim was to evaluate the performance of machine learning models using admission data to predict the likely length of stay (LOS) for patients admitted with stroke. Secondary aims included the prediction of discharge modified Rankin Scale (mRS), in-hospital mortality, and discharge destination. In this study a retrospective dataset was used to develop and test a variety of machine learning models. The patients included in the study were all stroke admissions (both ischaemic stroke and intracerebral haemorrhage) at a single tertiary hospital between December 2016 and September 2019. The machine learning models developed and tested (75%/25% train/test split) included logistic regression, random forests, decision trees and artificial neural networks. The study included 2840 patients. In LOS prediction the highest area under the receiver operator curve (AUC) was achieved on the unseen test dataset by an artificial neural network at 0.67. Higher AUC were achieved using logistic regression models in the prediction of discharge functional independence (mRS ≤2) (AUC 0.90) and in the prediction of in-hospital mortality (AUC 0.90). Logistic regression was also the best performing model for predicting home vs non-home discharge destination (AUC 0.81). This study indicates that machine learning may aid in the prognostication of factors relevant to post-stroke discharge planning. Further prospective and external validation is required, as well as assessment of the impact of subsequent implementation.

A semiautomatic segmentation method for interstitial needles in intraoperative 3D transvaginal ultrasound images for high-dose-rate gynecologic brachytherapy of vaginal tumors

PurposeThe purpose of this study was to evaluate the use of a semiautomatic algorithm to simultaneously segment multiple high-dose-rate (HDR) gynecologic interstitial brachytherapy (ISBT) needles in three-dimensional (3D) transvaginal ultrasound (TVUS) images, with the aim of providing a clinically useful tool for intraoperative implant assessment. Methods and MaterialsA needle segmentation algorithm previously developed for HDR prostate brachytherapy was adapted and extended to 3D TVUS images from gynecologic ISBT patients with vaginal tumors. Two patients were used for refining/validating the modified algorithm and five patients (8–12 needles/patient) were reserved as an unseen test data set. The images were filtered to enhance needle edges, using intensity peaks to generate feature points, and leveraged the randomized 3D Hough transform to identify candidate needle trajectories. Algorithmic segmentations were compared against manual segmentations and calculated dwell positions were evaluated. ResultsAll 50 test data set needles were successfully segmented with 96% of algorithmically segmented needles having angular differences <3° compared with manually segmented needles and the maximum Euclidean distance was <2.1 mm. The median distance between corresponding dwell positions was 0.77 mm with 86% of needles having maximum differences <3 mm. The mean segmentation time using the algorithm was <30 s/patient. ConclusionsWe successfully segmented multiple needles simultaneously in intraoperative 3D TVUS images from gynecologic HDR-ISBT patients with vaginal tumors and demonstrated the robustness of the algorithmic approach to image artifacts. This method provided accurate segmentations within a clinically efficient timeframe, providing the potential to be translated into intraoperative clinical use for implant assessment.

A CT perfusion based model predicts outcome in wake-up stroke patients treated with recombinant tissue plasminogen activator

Objective: Advanced neuroimaging has proved to be pivotal in the management of acute ischemic stroke. The use of CT perfusion (CTP) core and penumbra parameters to predict the outcome in wake-up stroke (WUS) patients in everyday clinical scenarios has not yet been investigated. The aim of our study was to investigate the predictive power of CTP parameters on functional and morphological outcomes in WUS patients treated with recombinant tissue plasminogen activator (rTPA). Approach: We analyzed clinical data and processed CTP images of 83 consecutive WUS patients treated with rTPA. The predictive power of whole-brain CTP features and of the clinical stroke-related parameters to predict the National Institutes of Health Stroke Scale (NIHSS) score at the seventh day and ischemic lesion volume outcome was investigated by means of multivariate regression analysis as well as least absolute shrinkage and selection operator (LASSO) modeling. Main results: Multivariate analysis showed that CTP core volume (β = 0.403, p = 0.000), NIHSS at admission (β = 0.323, p = 0.005) and Alberta Stroke Program Early CT (ASPECT) score (β = −0.224, p = 0.012) predict NIHSS at 7 days, while total hypoperfused volume (β = 0.542, p = 0.000) and core volume on CTP (β =0.441, p = 0.000) predict infarct lesion volume at follow-up CT. The LASSO modeling approach confirmed the significant predictive power of CTP core volume, total hypoperfused CTP volume, NIHSS at baseline and ASPECT score, producing a sparse model with adequate reliability (the root mean square error on a previously unseen testing dataset was 3.68). Significance: Our findings highlight the importance of CT multimodal imaging features for decision-making and prediction in the hyperacute phase of WUS. The predictive model supports the hypothesis that an irreversible necrotic core rather than the extent of the penumbra is the main prognostic factor in WUS patients treated with rTPA.

Predicting motor outcome in preterm infants from very early brain diffusion MRI using a deep learning convolutional neural network (CNN) model

Background and aimsPreterm birth imposes a high risk for developing neuromotor delay. Earlier prediction of adverse outcome in preterm infants is crucial for referral to earlier intervention. This study aimed to predict abnormal motor outcome at 2 years from early brain diffusion magnetic resonance imaging (MRI) acquired between 29 and 35 weeks postmenstrual age (PMA) using a deep learning convolutional neural network (CNN) model. MethodsSeventy-seven very preterm infants (born <31 weeks gestational age (GA)) in a prospective longitudinal cohort underwent diffusion MR imaging (3T Siemens Trio; 64 directions, b ​= ​2000 ​s/mm2). Motor outcome at 2 years corrected age (CA) was measured by Neuro-Sensory Motor Developmental Assessment (NSMDA). Scores were dichotomised into normal (functional score: 0, normal; n ​= ​48) and abnormal scores (functional score: 1–5, mild-profound; n ​= ​29). MRIs were pre-processed to reduce artefacts, upsampled to 1.25 ​mm isotropic resolution and maps of fractional anisotropy (FA) were estimated. Patches extracted from each image were used as inputs to train a CNN, wherein each image patch predicted either normal or abnormal outcome. In a postprocessing step, an image was classified as predicting abnormal outcome if at least 27% (determined by a grid search to maximise the model performance) of its patches predicted abnormal outcome. Otherwise, it was considered as normal. Ten-fold cross-validation was used to estimate performance. Finally, heatmaps of model predictions for patches in abnormal scans were generated to explore the locations associated with abnormal outcome. ResultsFor the identification of infants with abnormal motor outcome based on the FA data from early MRI, we achieved mean sensitivity 70% (standard deviation SD 19%), mean specificity 74% (SD 39%), mean AUC (area under the receiver operating characteristic curve) 72% (SD 14%), mean F1 score of 68% (SD 13%) and mean accuracy 73% (SD 19%) on an unseen test data set. Patch-based prediction heatmaps showed that the patches around the motor cortex and somatosensory regions were most frequently identified by the model with high precision (74%) as a location associated with abnormal outcome. Part of the cerebellum, and occipital and frontal lobes were also highly associated with abnormal NSMDA/motor outcome. Discussion/conclusionThis study established the potential of an early brain MRI-based deep learning CNN model to identify preterm infants at risk of a later motor impairment and to identify brain regions predictive of adverse outcome. Results suggest that predictions can be made from FA maps of diffusion MRIs well before term equivalent age (TEA) without any prior knowledge of which MRI features to extract and associated feature extraction steps. This method, therefore, is suitable for any case of brain condition/abnormality. Future studies should be conducted on a larger cohort to re-validate the robustness and effectiveness of these models.

How Well Can the Migration Component of Regional Population Change be Predicted? A Machine Learning Approach Applied to German Municipalities

For several decades, demographic forecasts had predicted that the majority of Germany’s cities and rural areas would experience population decline in the early 21st century. Instead, recent trends show a growing population size in three out of every four German districts. As a result, there are currently severe shortages of housing and childcare in regions that were projected to decline but have instead grown in recent years. Other regions, by contrast, continue to lose young people in particular. Most of these differences between regions stem from within-country as well as international migration. An important question for both regional demographic research as well as local policy-makers is thus how well net migration rates in cities and rural districts can be predicted several years into the future. In this study, we develop models that predict migration (both within-country as well as international migration) at the level of municipalities for two demographic groups, namely young people aged 18 to 24 years, and families (people aged 30 to 49 years and underage children). We collect data on economic, demographic and other characteristics such as distances to large cities or universities for around 3,000 German municipalities (Gemeinden). The model is trained on a subset of these data from the period 2005-2009 and predicts net migration rates among young people on an unseen test dataset in the future (i.e. for the period 2011-2015). The results show that the model can predict future net migration by young people aged 18 to 24 years reasonably well (R² &gt; 0.5), although there were quite significant changes during the period under study, for example refugee immigration to Germany. Family migration, on the other hand, cannot be predicted equally well (R² = 0.25). Some important lessons emerge concerning the predictability of regional and international migration and the usefulness of demographic forecasts for local policy-makers.

Modeling high-temperature mechanical properties of austenitic stainless steels by neural networks

An artificial neural network (ANN) model was designed to correlate the complex relations among composition, temperature, and mechanical properties of 18Cr-12Ni-Mo austenitic stainless steels. The developed model was used to estimate the composition-property and temperature-property correlations with 97% and 91% accuracy, for train and unseen test datasets. The ANN predictions are more accurate with experimental results as compared with the calculated properties of the existing model. The effective response of the alloying elements on the mechanical properties at ambient as well as elevated temperatures was quantitatively estimated with the help of the index of relative importance (IRI). The calculated results of the ANN model beneficial for both researchers as well as designers to guide actual experiments. Hence, this proposed technique will be helpful in developing the components of austenitic stainless steel with desired properties.

Three-Plane-assembled Deep Learning Segmentation of Gliomas.

To design a computational method for automatic brain glioma segmentation of multimodal MRI scans with high efficiency and accuracy. The 2018 Multimodal Brain Tumor Segmentation Challenge (BraTS) dataset was used in this study, consisting of routine clinically acquired preoperative multimodal MRI scans. Three subregions of glioma-the necrotic and nonenhancing tumor core, the peritumoral edema, and the contrast-enhancing tumor-were manually labeled by experienced radiologists. Two-dimensional U-Net models were built using a three-plane-assembled approach to segment three subregions individually (three-region model) or to segment only the whole tumor (WT) region (WT-only model). The term three-plane-assembled means that coronal and sagittal images were generated by reformatting the original axial images. The model performance for each case was evaluated in three classes: enhancing tumor (ET), tumor core (TC), and WT. On the internal unseen testing dataset split from the 2018 BraTS training dataset, the proposed models achieved mean Sørensen-Dice scores of 0.80, 0.84, and 0.91, respectively, for ET, TC, and WT. On the BraTS validation dataset, the proposed models achieved mean 95% Hausdorff distances of 3.1 mm, 7.0 mm, and 5.0 mm, respectively, for ET, TC, and WT and mean Sørensen-Dice scores of 0.80, 0.83, and 0.91, respectively, for ET, TC, and WT. On the BraTS testing dataset, the proposed models ranked fourth out of 61 teams. The source code is available at https://github.com/GuanLab/Brain_Glioma. This deep learning method consistently segmented subregions of brain glioma with high accuracy, efficiency, reliability, and generalization ability on screening images from a large population, and it can be efficiently implemented in clinical practice to assist neuro-oncologists or radiologists. Supplemental material is available for this article. © RSNA, 2020.

Quantitative Thermal Imaging Biomarkers to Detect Acute Skin Toxicity From Breast Radiation Therapy Using Supervised Machine Learning

Radiation-induced dermatitis is a common side effect of breast radiation therapy (RT). Current methods to evaluate breast skin toxicity include clinical examination, visual inspection, and patient-reported symptoms. Physiological changes associated with radiation-induced dermatitis, such as inflammation, may also increase body-surface temperature, which can be detected by thermal imaging. Quantitative thermal imaging markers were identified and used in supervised machine learning to develop a predictive model for radiation dermatitis. Ninety patients treated for adjuvant whole-breast RT (4250 cGy/fx = 16) were recruited for the study. Thermal images of the treated breast were taken at 4 intervals: before RT, then weekly at fx = 5, fx = 10, and fx= 15. Parametric thermograms were analyzed and yielded 26 thermal-based features that included surface temperature (°C) and texture parameters obtained from (1) gray-level co-occurrence matrix, (2) gray-level run-length matrix, and (3) neighborhood gray-tone difference matrix. Skin toxicity was evaluated at the end of RT using the Common Terminology Criteria for Adverse Events (CTCAE) guidelines (Ver.5). Binary group classes were labeled according to a CTCAE cut-off score of ≥2, and thermal features obtained at fx = 5 were used for supervised machine learning to predict skin toxicity. The data set was partitioned for model training, independent testing, and validation. Fifteen patients (∼17% of the whole data set) were randomly selected as an unseen test data set, and 75 patients (∼83% of the whole data set) were used for training and validation of the model. A random forest classifier with leave-1-patient-out cross-validation was employed for modeling single and hybrid parameters. The model performance was reported using receiver operating characteristic analysis on patients from an independent test set. Thirty-seven patients presented with adverse skin effects, denoted by a CTCAE score ≥2, and had significantly higher local increases in skin temperature, reaching 36.06°C at fx = 10 (P = .029). However, machine-learning models demonstrated early thermal signals associated with skin toxicity after the fifth RT fraction. The cross-validated model showed high prediction accuracy on the independent test data (test accuracy = 0.87) at fx = 5 for predicting skin toxicity at the end of RT. Early thermal markers after 5 fractions of RT are predictive of radiation-induced skin toxicity in breast RT.

Application of radiomics for the prediction of HPV status for patients with head and neck cancers.

To perform radiomic analysis of primary tumors extracted from pretreatment contrast-enhanced computed tomography (CE-CT) images for patients with oropharyngeal cancers to identify discriminant features and construct an optimal classifier for the characterization and prediction of human papilloma virus (HPV) status. One hundred and eighty seven patients with oropharyngeal cancers with known HPV status (confirmed by immunohistochemistry-p16 protein testing) were retrospectively studied as follows: Group A: 95 patients (19HPV- and 76HPV+) from the MICAII grand challenge. Group B: 92 patients (52HPV- and 40HPV+) from our institution. Radiomic features (172) were extracted from pretreatment diagnostic CE-CT images of the gross tumor volume (GTV). Levene and Kolmogorov-Smirnov's tests with absolute biserial correlation (>0.48) were used to identify the discriminant features between the HPV+ and HPV- groups. The discriminant features were used to train and test eight different classifiers. Area under receiver operating characteristic (AUC), positive predictive and negative predictive values (PPV and NPV, respectively) were used to evaluate the performance of the classifiers. Principal component analysis (PCA) was applied on the discriminant feature set and seven PCs were used to train and test a generalized linear model (GLM) classifier. Among 172 radiomic features only 12 radiomic features (from 3 categories) were significantly different (P<0.05, |BSC|>0.48) between the HPV+ and HPV- groups. Among the eight classifiers trained and applied for prediction of HPV status, the GLM showed the highest performance for each discriminant feature and the combined 12 features: AUC/PPV/NPV=0.878/0.834/0.811. The GLM high prediction power was AUC/PPV/NPV=0.849/0.731/0.788 and AUC/PPV/NPV=0.869/0.807/0.870 for unseen test datasets for groups A and B, respectively. After eliminating the correlation among discriminant features by applying PCA analysis, the performance of the GLM was improved by 3.3%, 2.2%, and 1.8% for AUC, PPV, and NPV, respectively. Results imply that GTV's for HPV+ patients exhibit higher intensities, smaller lesion size, greater sphericity/roundness, and higher spatial intensity-variation/heterogeneity. Results are suggestive that radiomic features primarily associated with the spatial arrangement and morphological appearance of the tumor on contrast-enhanced diagnostic CT datasets may be potentially used for classification of HPV status.

MapReduce-Based Parallel Genetic Algorithm for CpG-Site Selection in Age Prediction.

Genomic biomarkers such as DNA methylation (DNAm) are employed for age prediction. In recent years, several studies have suggested the association between changes in DNAm and its effect on human age. The high dimensional nature of this type of data significantly increases the execution time of modeling algorithms. To mitigate this problem, we propose a two-stage parallel algorithm for selection of age related CpG-sites. The algorithm first attempts to cluster the data into similar age ranges. In the next stage, a parallel genetic algorithm (PGA), based on the MapReduce paradigm (MR-based PGA), is used for selecting age-related features of each individual age range. In the proposed method, the execution of the algorithm for each age range (data parallel), the evaluation of chromosomes (task parallel) and the calculation of the fitness function (data parallel) are performed using a novel parallel framework. In this paper, we consider 16 different healthy DNAm datasets that are related to the human blood tissue and that contain the relevant age information. These datasets are combined into a single unioned set, which is in turn randomly divided into two sets of train and test data with a ratio of 7:3, respectively. We build a Gradient Boosting Regressor (GBR) model on the selected CpG-sites from the train set. To evaluate the model accuracy, we compared our results with state-of-the-art approaches that used these datasets, and observed that our method performs better on the unseen test dataset with a Mean Absolute Deviation (MAD) of 3.62 years, and a correlation (R2) of 95.96% between age and DNAm. In the train data, the MAD and R2 are 1.27 years and 99.27%, respectively. Finally, we evaluate our method in terms of the effect of parallelization in computation time. The algorithm without parallelization requires 4123 min to complete, whereas the parallelized execution on 3 computing machines having 32 processing cores each, only takes a total of 58 min. This shows that our proposed algorithm is both efficient and scalable.