Classes Of Features Research Articles

Biomechanical Effects of Thoracolumbosacral Orthosis Design Features on 3D Correction in Adolescent Idiopathic Scoliosis: A Comprehensive Multicenter Study.

Multicenter numerical study. To biomechanically analyze and compare various passive correction features of braces, designed by several centers with diverse practices, for three-dimensional (3D) correction of adolescent idiopathic scoliosis. A wide variety of brace designs exist, but their biomechanical effectiveness is not clearly understood. Many studies have reported brace treatment correction potential with various degrees of control, making the objective comparison of correction mechanisms difficult. A Finite Element Model simulating the immediate in-brace corrective effects has been developed and allows to comprehensively assess the biomechanics of different brace designs. Expert clinical teams (one orthotist and one orthopedist) from six centers in five countries participated in the study. For six scoliosis cases with different curve types respecting SRS criteria, the teams designed two braces according to their treatment protocol. Finite Element Model simulations were performed to compute immediate in-brace 3D correction and skin-to-brace pressures. All braces were randomized and labeled according to 21 design features derived from Society on Scoliosis Orthopaedic and Rehabilitation Treatment proposed descriptors, including positioning of pressure points, orientation of push vectors, and sagittal design. Simulated in brace 3D corrections were compared for each design feature class using ANOVAs and linear regressions (significance P ≤ 0.05). Seventy-two braces were tested, with significant variety in the design approaches. Pressure points at the apical vertebra level corrected the main thoracic curve better than more caudal locations. Braces with ventral support flattened the lumbar lordosis. Lateral and ventral skin-to-brace pressures were correlated with changes in thoracolumbar/lumbar Cobb and lumbar lordosis (r =- 0.53, r = - 0.54). Upper straps positioned above T10 corrected the main thoracic Cobb better than those placed lower. The corrective features of various scoliosis braces were objectively compared in a systematic approach with minimal biases and variability in test parameters, providing a better biomechanical understanding of individual passive mechanisms' contribution to 3D correction.

Selection of Algorithm for Land Use Land Cover Classification and Change Detection

Land use land cover (LULC) classification is an important research criterion for urban growth modelling. In this study, five different parametric classification algorithms (Maximum Likelihood Classifier, Mahalanobis Distance, Minimum Distance, Spectral Angle Mapper, Spectral Correlation Mapper) are used for the LULC classification of the Barrackpore subdivision of West Bengal state in India. Two different Landsat datasets (Landsat5 and Landsat8) are used for 2005 and 2020. Various algorithms have shown good accuracy for different LULC feature classes. Minimum Distance and Maximum Likelihood Classifier (MLC) has given the highest overall accuracy in 2005 and 2020, respectively. The best classification results are used for the change detection in the LULC classes over fifteen years (2005-2020). The classification result will help to choose a suitable algorithm for a regional level LULC study in further research. The change detection result indicates the need for a good growth pattern in the region.

Exploring the Utility of Radiomic Feature Extraction to Improve the Diagnostic Accuracy of Cardiac Sarcoidosis Using FDG PET.

BackgroundThis study aimed to explore the radiomic features from PET images to detect active cardiac sarcoidosis (CS).MethodsForty sarcoid patients and twenty-nine controls were scanned using FDG PET-CMR. Five feature classes were compared between the groups. From the PET images alone, two different segmentations were drawn. For segmentation A, a region of interest (ROI) was manually delineated for the patients' myocardium hot regions with standardized uptake value (SUV) higher than 2.5 and the controls' normal myocardium region. A second ROI was drawn in the entire left ventricular myocardium for both study groups, segmentation B. The conventional metrics and radiomic features were then extracted for each ROI. Mann-Whitney U-test and a logistic regression classifier were used to compare the individual features of the study groups.ResultsFor segmentation A, the SUVmin had the highest area under the curve (AUC) and greatest accuracy among the conventional metrics. However, for both segmentations, the AUC and accuracy of the TBRmax were relatively high, >0.85. Twenty-two (from segmentation A) and thirty-five (from segmentation B) of 75 radiomic features fulfilled the criteria: P-value < 0.00061 (after Bonferroni correction), AUC >0.5, and accuracy >0.7. Principal Component Analysis (PCA) was conducted, with five components leading to cumulative variance higher than 90%. Ten machine learning classifiers were then tested and trained. Most of them had AUCs and accuracies ≥0.8. For segmentation A, the AUCs and accuracies of all classifiers are >0.9, but k-neighbors and neural network classifiers were the highest (=1). For segmentation B, there are four classifiers with AUCs and accuracies ≥0.8. However, the gaussian process classifier indicated the highest AUC and accuracy (0.9 and 0.8, respectively).ConclusionsRadiomic analysis of the specific PET data was not proven to be necessary for the detection of CS. However, building an automated procedure will help to accelerate the analysis and potentially lead to more reproducible findings across different scanners and imaging centers and consequently improve standardization procedures that are important for clinical trials and development of more robust diagnostic protocols.

Prediction of Potentially Suitable Distribution Areas for Prunus tomentosa in China Based on an Optimized MaxEnt Model

Prunus tomentosa (Thunb.) Wall has high nutritional value and medicinal effects. It is widespread in China; however, most plants growing in the wild are near extinction in many places. Predicting the potential distribution of P. tomentosa under climate change is helpful for cultivating and protecting wild germplasm resources. We used two general circulation models (CCSM4 and MIROC-ESM) and two future climate scenarios (RCP4.5 and RCP8.5) to predict P. tomentosa’s present and future geographical distribution. A total of 137 distribution data points and 19 bioclimatic variables were imported into the maximum entropy model (MaxEnt). The optimal parameter combination (feature class LQHPT, regularized multiplier 3.0) was selected with corrected Akaike Information Criterion as the index. The results showed that at present and in the future, P. tomentosa was distributed across the northern provinces, with Gansu, Shanxi, Shaanxi, and Henan being the most suitable regions. Compared with the current climatic conditions, the potential growing area of P. tomentosa will move north, and the growing area will increase, especially in Xinjiang, where the low-impact zone area decreases. Temperature and humidity were the main variables affecting the potential distribution of the plant, including the average temperature in the coldest season (Bio11) and precipitation in the warmest season (Bio18).

Impact of ComBat Harmonization on PET Radiomics-Based Tissue Classification: A Dual-Center PET/MRI and PET/CT Study.

Our purpose was to determine whether ComBat harmonization improves 18F-FDG PET radiomics-based tissue classification in pooled PET/MRI and PET/CT datasets. Methods: Two hundred patients who had undergone 18F-FDG PET/MRI (2 scanners and vendors; 50 patients each) or PET/CT (2 scanners and vendors; 50 patients each) were retrospectively included. Gray-level histogram, gray-level cooccurrence matrix, gray-level run-length matrix, gray-level size-zone matrix, and neighborhood gray-tone difference matrix radiomic features were calculated for volumes of interest in the disease-free liver, spleen, and bone marrow. For individual feature classes and a multiclass radiomic signature, tissue was classified on ComBat-harmonized and unharmonized pooled data, using a multilayer perceptron neural network. Results: Median accuracies in training and validation datasets were 69.5% and 68.3% (harmonized), respectively, versus 59.5% and 58.9% (unharmonized), respectively, for gray-level histogram; 92.1% and 86.1% (harmonized), respectively, versus 53.6% and 50.0% (unharmonized), respectively, for gray-level cooccurrence matrix; 84.8% and 82.8% (harmonized), respectively, versus 62.4% and 58.3% (unharmonized), respectively, for gray-level run-length matrix; 87.6% and 85.6% (harmonized), respectively, versus 56.2% and 52.8% (unharmonized), respectively, for gray-level size-zone matrix; 79.5% and 77.2% (harmonized), respectively, versus 54.8% and 53.9% (unharmonized), respectively, for neighborhood gray-tone difference matrix; and 86.9% and 84.4% (harmonized), respectively, versus 62.9% and 58.3% (unharmonized), respectively, for radiomic signature. Conclusion: ComBat harmonization may be useful for multicenter 18F-FDG PET radiomics studies using pooled PET/MRI and PET/CT data.

Feature-dependent group structures and hierarchical songbird-habitat relationships in a managed forest landscape

We examined the importance of local habitat features on forest songbird distributions using a feature-dependent group framework that organizes species in relation to variable importance of species-habitat relationships. Habitat features included vegetation structure, floristics, and environmental (i.e., topography) classes. We also examined evidence for hierarchical habitat selection by assessing feature positioning within the ensemble structural models, Random Forests. From the 44 species examined, we distinguished 14 groups, comprising two to eight species per group, and examined variation in habitat importance across the groups. Overall, floristic species turnover, shrub density, canopy height, and tree basal area were among the most important features. On average, however, no specific feature class was more important than the others across the entire collection of species. Instead, we interpreted group partitioning to occur primarily by species differentiating among specific forest types, such as bottomland riparian or oak-dominated (Quercus spp.) stands, and species responding to structural variation across successional gradients. The positioning of habitat features within individual tree models showed that for those species among specific forest types, floristics features were selected first in the models followed by environmental features. Whereas when vegetation structural features were most important, structural features were typically selected first followed by floristics. The large variability in responses likely reflects the large heterogeneity in vegetation structure and composition present in our Central Hardwood, USA, forest landscape. In such cases, avian community assembly processes incorporated the primary moisture (mesic-xeric tree species) and successional (open- to closed-canopy) vegetative gradients of these forestlands, which underscores the importance of maintaining vegetation heterogeneity under accelerating environmental change. By including diverse suites of habitat types, we demonstrate how these features vary in their importance across songbird assemblages in a diverse forestland and identify key habitat features in the conservation of these bird communities.

Multi-omics data integration reveals metabolome as the top predictor of the cervicovaginal microenvironment.

Emerging evidence suggests that host-microbe interaction in the cervicovaginal microenvironment contributes to cervical carcinogenesis, yet dissecting these complex interactions is challenging. Herein, we performed an integrated analysis of multiple “omics” datasets to develop predictive models of the cervicovaginal microenvironment and identify characteristic features of vaginal microbiome, genital inflammation and disease status. Microbiomes, vaginal pH, immunoproteomes and metabolomes were measured in cervicovaginal specimens collected from a cohort (n = 72) of Arizonan women with or without cervical neoplasm. Multi-omics integration methods, including neural networks (mmvec) and Random Forest supervised learning, were utilized to explore potential interactions and develop predictive models. Our integrated analyses revealed that immune and cancer biomarker concentrations were reliably predicted by Random Forest regressors trained on microbial and metabolic features, suggesting close correspondence between the vaginal microbiome, metabolome, and genital inflammation involved in cervical carcinogenesis. Furthermore, we show that features of the microbiome and host microenvironment, including metabolites, microbial taxa, and immune biomarkers are predictive of genital inflammation status, but only weakly to moderately predictive of cervical neoplastic disease status. Different feature classes were important for prediction of different phenotypes. Lipids (e.g. sphingolipids and long-chain unsaturated fatty acids) were strong predictors of genital inflammation, whereas predictions of vaginal microbiota and vaginal pH relied mostly on alterations in amino acid metabolism. Finally, we identified key immune biomarkers associated with the vaginal microbiota composition and vaginal pH (MIF), as well as genital inflammation (IL-6, IL-10, MIP-1α).

Feature reduction for imbalanced data classification using similarity-based feature clustering with adaptive weighted K-nearest neighbors

Most existing imbalanced data classification models mainly focus on the classification performance of majority class samples, and many clustering algorithms need to manually specify the initial cluster centers and the number of clusters. To solve these drawbacks, this study presents a novel feature reduction method for imbalanced data classification using similarity-based feature clustering with adaptive weighted k-nearest neighbors (AWKNN). First, the similarity between samples is evaluated by the difference and smaller value between samples on each dimension, a similarity measure matrix is then developed to measure the similarity between clusters, after which a new hierarchical clustering model is constructed. By combining the cluster center of each sample cluster with its nearest neighbor, new samples are generated. Then, a hybrid sampling model based on similarity measure is presented by putting the generated samples into imbalanced data and removing samples from majority classes. Thus, a balanced decision system is constructed based on generated samples and minority class samples. Second, to address the issues that the traditional symmetric uncertainty only considers the correlation between features, and mutual information ignores the added information after classification, the normalized information gain is introduced to design new symmetric uncertainty between each feature and the other features; then, the ordered sequence and the average of the symmetric uncertainty difference of each feature are provided to adaptively select the k-nearest neighbors of features. Moreover, the weight of the k-th nearest neighbor of features is defined to present the AWKNN density of features and their ordered sequence for clustering features. Finally, by combining the weighted average redundancy with the symmetric uncertainty between features and decision classes, the maximum relevance between each feature and decision classes, and the minimum redundancy among features in the same cluster is presented to select the optimal feature subset from the feature clusters. Experiments applied to 29 imbalanced datasets show that the developed algorithm is effective and can select the optimal feature subset with high classification accuracy for imbalanced data.

A Machine Learning Method for Classification of Cervical Cancer

Cervical cancer is one of the leading causes of premature mortality among women worldwide and more than 85% of these deaths are in developing countries. There are several risk factors associated with cervical cancer. In this paper, we developed a predictive model for predicting the outcome of patients with cervical cancer, given risk patterns from individual medical records and preliminary screening. This work presents a decision tree (DT) classification algorithm to analyze the risk factors of cervical cancer. Recursive feature elimination (RFE) and least absolute shrinkage and selection operator (LASSO) feature selection techniques were fully explored to determine the most important attributes for cervical cancer prediction. The dataset employed here contains missing values and is highly imbalanced. Therefore, a combination of under and oversampling techniques called SMOTETomek was employed. A comparative analysis of the proposed model has been performed to show the effectiveness of feature selection and class imbalance based on the classifier’s accuracy, sensitivity, and specificity. The DT with the selected features from RFE and SMOTETomek has better results with an accuracy of 98.72% and sensitivity of 100%. DT classifier is shown to have better performance in handling classification problems when the features are reduced, and the problem of high class imbalance is addressed.

Multi‐label hybrid radar signal recognition based on a feature pyramid network and class activation mapping

The electromagnetic environment of modern battlefields becomes increasingly complex, and radar receivers may receive multiple radar signals simultaneously. However, current deep learning models can only predict a single class and cannot recognize multi-label mixed radar signals. In this study, a multi-label hybrid radar signal recognition framework based on the feature pyramid network (FPN) and class activation map (CAM) is proposed. The multi-label radar signals are recognized by calculating the average value of the CAM corresponding to each class. The proposed method can recognize, localize and separate mixed radar signals in time-frequency images, which improves the interpretability and transparency of the model. In addition, the FPN is adopted to improve the spatial resolution of the feature maps, and the Mixup data augmentation is utilized to improve the generalization performance of the model. Experiments with eight different modulation types of mixed radar signals show that the recognition accuracy of hybrid radar signals achieves 92.2% at 0 dB.

Scale-adaptive local intentional surface feature detection

In this paper, we propose a mesh-based feature detection scheme that focuses on surface features. A class of features of key interest is intentional structures that act as fiducials and that, for instance, can assist in shape retrieval and distortion measurement. We introduce a tunable two-scale depth measurement scheme to quantify the displacement of a vertex from the local surface, which can be a strong indicator of features. We print and scan 3D models with fiducial features appearing across the surface to demonstrate the high fidelity and accuracy of the proposed feature detection scheme. The method outperforms existing 3D feature detection schemes on CAD models and 3D scans alike. We also discuss applications of data embedding enabled by the achievable detection performance.

Comparison of the Novel Probabilistic Self-Optimizing Vectorized Earth Observation Retrieval Classifier with Common Machine Learning Algorithms

The Vectorized Earth Observation Retrieval (VEOR) algorithm is a novel algorithm suited to the efficient supervised classification of large Earth Observation (EO) datasets. VEOR addresses shortcomings in well-established machine learning methods with an emphasis on numerical performance. Its characteristics include (1) derivation of classification probability; (2) objective selection of classification features that maximize Cohen’s kappa coefficient (κ) derived from iterative “leave-one-out” cross-validation; (3) reduced sensitivity of the classification results to imbalanced classes; (4) smoothing of the classification probability field to reduce noise/mislabeling; (5) numerically efficient retrieval based on a pre-computed look-up vector (LUV); and (6) separate parametrization of the algorithm for each discrete feature class (e.g., land cover). Within this study, the performance of the VEOR classifier was compared to other commonly used machine learning algorithms: K-nearest neighbors, support vector machines, Gaussian process, decision trees, random forest, artificial neural networks, AdaBoost, Naive Bayes and Quadratic Discriminant Analysis. Firstly, the comparison was performed using synthetic 2D (two-dimensional) datasets featuring different sample sizes, levels of noise (i.e., mislabeling) and class imbalance. Secondly, the same experiments were repeated for 7D datasets consisting of informative, redundant and insignificant features. Ultimately, the benchmarking of the classifiers involved cloud discrimination using MODIS satellite spectral measurements and a reference cloud mask derived from combined CALIOP lidar and CPR radar data. The results revealed that the proposed VEOR algorithm accurately discriminated cloud cover using MODIS data and accurately classified large synthetic datasets with low or moderate levels of noise and class imbalance. On the contrary, VEOR did not feature good classification skills for significantly distorted or for small datasets. Nevertheless, the comparisons performed proved that VEOR was within the 3–4 most accurate classifiers and that it can be applied to large Earth Observation datasets.

Neighborhood Rough Set Approach With Biometric Application

This paper provides a new approach for human identification based on Neighborhood Rough Set (NRS) algorithm with biometric application of ear recognition. The traditional rough set model can just be used to evaluate categorical features. The neighborhood model is used to evaluate both numerical and categorical features by assigning different thresholds for different classes of features. The feature vectors are obtained from ear image and ear matching process is performed. Actually, matching is a process of ear identification. The extracted features are matched with classes of ear images enrolled in the database. NRS algorithm is developed in this work for feature matching. A set of 20 persons are used for experimental analysis and each person is having six images. The experimental result illustrates the high accuracy of NRS approach when compared to other existing techniques.

Fine-Grained Sediment and Phosphorus Storage in a Suspended-Load-Dominated, Alluvial Channel

HighlightsIn-channel sediment and TP storage were quantified along 13.5 km of a stream.TP and sediment masses were 1 and 3 times greater than their respective annual watershed load.Colluvial material at streambank toe represented the dominant storage pool.Sinuous stream reaches exhibited significantly greater storage versus channelized reaches.Abstract. In-channel storage of both fine sediment and sediment-bound phosphorus has been recognized as potentially large contributors to respective watershed loads, yet they are rarely quantified in the field. In this field-based study, we quantified and characterized in-channel fine sediment and total phosphorus (TP) storage within a third-order stream. We hypothesized storage to be significant within the context of watershed-scale sediment and TP loads and that storage trends are influenced by channel characteristics and hydraulics. In May 2015, the volume of stored in-channel sediment was estimated within a 13.5 km stretch of Walnut Creek, an alluvial stream draining an agriculturally dominated watershed in Jasper County, Iowa, USA. Storage volume was estimated in-field through a series of 240 transects which were stratified across reaches of the varying channel and hydraulic conditions (e.g., sinuosity, stream power). Overall storage volume was broken down into feature classes (e.g., point bars) based on in-channel depositional processes. Following in-field quantification, feature class sediment samples were collected and analyzed for physical properties (e.g., bulk density) and TP concentration. Physical and chemical analyses, coupled with in-field volume estimation, allowed for watershed-scale masses of in-channel sediment and TP to be calculated and then placed within the context of respective annual loads. Sediment and TP were estimated to be stored at ~2.7 Mg m-1 and 0.7 kg m-1, respectively. Sinuous reaches exhibited significantly greater sediment storage volume (p &amp;lt; 0.001) and depth (p &amp;lt; 0.001) versus straight reaches. Reach sinuosity exhibited a significant positive correlation (p = 0.03) with storage mass and represented the most effective storage predictor. The majority of storage mass (72%) was represented by colluvial material accumulations at the streambank toe. Loose bed sediment was the second greatest contributor to storage mass (18%), with the remaining feature classes (e.g., bars) representing a combined ~10%. Sediment storage mass was ~3.25 times greater than the watershed suspended sediment load for 2015. The TP mass was found to be nearly equal to the respective watershed load for 2015. In-channel sediment and TP storage masses within Walnut Creek were found to be significant in comparison to respective annual loads, and thus storage is expected to play a significant role in respective watershed routing, loading, and the overall budgets of these parameters. Though challenging, quantification of in-channel fine sediment and TP storage is necessary to relate source contributions (e.g., streambank erosion) to watershed loads and is critical to account for interannual sediment and TP export. Keywords: Channel evolution, Channel morphology, Iowa, Sinuosity, Stream power, Water quality, Watershed export.

Detection of Blob and Feature Envy Smells in a Class Diagram using Class's Features

Detection of Blob and Feature Envy Smells in a Class Diagram using Class's Features

Tolerance analysis of wedge assembly considering impact force

This paper proposes a dynamic theory based tolerance analysis method for wedge assembly. Wedge is a class of contact features with a small contact angle such as cone and triangle. The impact usually applied in wedge assembly to tighten the structure. The force and times of the impact could affect the assembly dimension and must be considered in tolerance analysis. The wedge contact is formulated as the dynamic model, which the model could be solved to obtain the variation length. Then the variation could be compensated to the tolerance model or dimension chain. The proposed method is used in a case of wedge belt assembly for verification.

Comprehensive DDoS Attack Classification Using Machine Learning Algorithms

The fast development of Internet technologies ignited the growth of techniques for information security that protect data, networks, systems, and applications from various threats. There are many types of threats. The dedicated denial of service attack (DDoS) is one of the most serious and widespread attacks on Internet resources. This attack is intended to paralyze the victim's system and cause the service to fail. This work is devoted to the classification of DDoS attacks in the special network environment called Software-Defined Networking (SDN) using machine learning algorithms. The analyzed dataset included instances of two classes: benign and malicious. As the dataset contained twenty-two features, the feature selection techniques were required for dimensionality reduction. In these experiments, the Information gain, the Chi-square, and the F-test were applied to decrease the number of features to ten. The classes were also not completely balanced, so undersampling, oversampling, and synthetic minority oversampling (SMOTE) techniques were used to balance classes equally. The previous research works observed the classification of DDoS attacks applying various feature selection techniques and one or more machine learning algorithms. Still, they did not pay much attention to classifying the combinations of feature selection and balancing methods with different machine learning algorithms. This work is devoted to the classification of datasets with eight machine learning algorithms: naïve Bayes, logistic regression, support vector machine, k-nearest neighbors, decision tree, random forest, XGBoost, and CatBoost. In the experimental results, the Information gain and F-test feature selection methods achieved better performance with all eight ML algorithms than with the Chi-square technique. Furthermore, the accuracy values of the oversampled and SMOTE datasets were higher than that of the undersampled and imbalanced datasets. Among machine learning algorithms, the accuracy of support vector machine, logistic regression, and naïve Bayes fluctuates between 0.59 and 0.75, while decision tree, random forest, XGBoost, and CatBoost allowed achieving values around 0.99 and 1.00 with all feature selection and class balancing techniques among all the algorithms.

Enhancing the Discovery of Functional Post-Translational Modification Sites with Machine Learning Models - Development, Validation, and Interpretation.

Protein posttranslational modifications (PTMs) are a rapidly expanding feature class of significant importance in cell biology. Due to a high burden of experimental proof, the number of functionals PTMs in the eukaryotic proteome is currently underestimated. Furthermore, not all PTMs are functionally equivalent. Computational approaches that can confidently recommend PTMs of probable function can improve the heuristics of PTM investigation and alleviate these problems. To address this need, we developed SAPH-ire: a multifeature heuristic neural network model that takes community wisdom into account by recommending experimental PTMs similar to those which have previously been established as having regulatory impact. Here, we describe the principle behind the SAPH-ire model, how it is developed, how we evaluate its performance, and important caveats to consider when building and interpreting such models. Finally, we discus current limitations of functional PTM prediction models and highlight potential mechanisms for their improvement.

Detecting Real-Time Fall of Elderly People Using Machine Learning

Abstract: Fall is a significant national health issue for the elderly people, generally resulting in severe injuries when the person lies down on the floor over an extended period without any aid after experiencing a great fall. Thus, elders need to be cared very attentively. A supervised-machine learning based fall detection approach with accelerometer, gyroscope is devised. The system can detect falls by grouping different actions as fall or non-fall events and the care taker is alerted immediately as soon as the person falls. The public dataset SisFall with efficient class of features is used to identify fall. The Random Forest (RF) and Support Vector Machine (SVM) machine learning algorithms are employed to detect falls with lesser false alarms. The SVM algorithm obtain a highest accuracy of 99.23% than RF algorithm. Keywords: Fall detection, Machine learning, Supervised classification, Sisfall, Activities of daily living, Wearable sensors, Random Forest, Support Vector Machine

A Comparative Study of Radiomics and Deep-Learning Based Methods for Pulmonary Nodule Malignancy Prediction in Low Dose CT Images.

ObjectivesBoth radiomics and deep learning methods have shown great promise in predicting lesion malignancy in various image-based oncology studies. However, it is still unclear which method to choose for a specific clinical problem given the access to the same amount of training data. In this study, we try to compare the performance of a series of carefully selected conventional radiomics methods, end-to-end deep learning models, and deep-feature based radiomics pipelines for pulmonary nodule malignancy prediction on an open database that consists of 1297 manually delineated lung nodules.MethodsConventional radiomics analysis was conducted by extracting standard handcrafted features from target nodule images. Several end-to-end deep classifier networks, including VGG, ResNet, DenseNet, and EfficientNet were employed to identify lung nodule malignancy as well. In addition to the baseline implementations, we also investigated the importance of feature selection and class balancing, as well as separating the features learned in the nodule target region and the background/context region. By pooling the radiomics and deep features together in a hybrid feature set, we investigated the compatibility of these two sets with respect to malignancy prediction.ResultsThe best baseline conventional radiomics model, deep learning model, and deep-feature based radiomics model achieved AUROC values (mean ± standard deviations) of 0.792 ± 0.025, 0.801 ± 0.018, and 0.817 ± 0.032, respectively through 5-fold cross-validation analyses. However, after trying out several optimization techniques, such as feature selection and data balancing, as well as adding context features, the corresponding best radiomics, end-to-end deep learning, and deep-feature based models achieved AUROC values of 0.921 ± 0.010, 0.824 ± 0.021, and 0.936 ± 0.011, respectively. We achieved the best prediction accuracy from the hybrid feature set (AUROC: 0.938 ± 0.010).ConclusionThe end-to-end deep-learning model outperforms conventional radiomics out of the box without much fine-tuning. On the other hand, fine-tuning the models lead to significant improvements in the prediction performance where the conventional and deep-feature based radiomics models achieved comparable results. The hybrid radiomics method seems to be the most promising model for lung nodule malignancy prediction in this comparative study.