Independent Test Dataset Research Articles

The effect of short-term training about depth predicting score on the diagnostic ability of invasion depth for differentiated early gastric Cancer among non-expert endoscopists

BackgroundThe depth-predicting score (DPS) was proposed based on conventional white-light imaging (C-WLI) endoscopic features of early gastric cancer (EGC) to determine the invasion depth of the neoplasm. However, the effect of DPS on training endoscopists remains unclear. Therefore, we aimed to investigate the effect of short-term DPS training on improving the diagnostic ability of EGC invasion depth and compare the training effect among non-expert endoscopists at different levels.MethodsIn the training session, the definitions and scoring rules of DPS were instructed, and classic C-WLI endoscopic example graphics were exhibited to the participants. Another C-WLI endoscopic images of 88 cases of histologically proven differentiated EGC were selected as an independent test dataset for evaluating the training effect. Each participant was tested, and the diagnostic accuracy rate of invasion depth was calculated differently one week before the training and after the completion of training.ResultsA total of 16 participants were enrolled and completed the training. Participants were divided into a trainee group and a junior endoscopist group according to the total number of C-WLI endoscopies performed. The total number of C-WLI endoscopies performed showed a significant difference between the trainee group and junior endoscopist group (350 vs. 2500, P = 0.001). No significant difference between the trainee group and junior endoscopist group was observed for pre-training accuracy. The overall diagnostic accuracy of invasion depth was improved significantly after completing DPS training compared with before (68.75 ± 5.71% vs. 61.58 ± 9.61%, P = 0.009). In the subgroup analysis, the post-training accuracy was higher than the pre-training accuracy, but significant improvement was observed only in the trainee group (61.65 ± 7.33% vs. 68.32 ± 5.71%, P = 0.034). In addition, no significant difference in post-training accuracy between the two groups was observed.ConclusionShort-term DPS training can improve the diagnostic ability of the invasion depth of EGC and homogenize the diagnostic ability of non-expert endoscopists at different levels. The depth-predicting score was convenient and effective for endoscopist training.

EmPDBA: protein-DNA binding affinity prediction by combining features from binding partners and interface learned with ensemble regression model.

Protein-deoxyribonucleic acid (DNA) interactions are important in a variety of biological processes. Accurately predicting protein-DNA binding affinity has been one of the most attractive and challenging issues in computational biology. However, the existing approaches still have much room for improvement. In this work, we propose an ensemble model for Protein-DNA Binding Affinity prediction (emPDBA), which combines six base models with one meta-model. The complexes are classified into four types based on the DNA structure (double-stranded or other forms) and the percentage of interface residues. For each type, emPDBA is trained with the sequence-based, structure-based and energy features from binding partners and complex structures. Through feature selection by the sequential forward selection method, it is found that there do exist considerable differences in the key factors contributing to intermolecular binding affinity. The complex classification is beneficial for the important feature extraction for binding affinity prediction. The performance comparison of our method with other peer ones on the independent testing dataset shows that emPDBA outperforms the state-of-the-art methods with the Pearson correlation coefficient of 0.53 and the mean absolute error of 1.11kcal/mol. The comprehensive results demonstrate that our method has a good performance for protein-DNA binding affinity prediction. Availability and implementation: The source code is available at https://github.com/ChunhuaLiLab/emPDBA/.

EpiTEAmDNA: Sequence feature representation via transfer learning and ensemble learning for identifying multiple DNA epigenetic modification types across species

Methylation is a major DNA epigenetic modification for regulating the biological processes without altering the DNA sequence, and multiple types of DNA methylations have been discovered, including 6mA, 5hmC, and 4mC. Multiple computational approaches were developed to automatically identify the DNA methylation residues using machine learning or deep learning algorithms. The machine learning (ML) based methods are difficult to be transferred to the other predicting tasks of the DNA methylation sites using additional knowledge. Deep learning (DL) may facilitate the transfer learning of knowledge from similar tasks, but they are often ineffective on small datasets. This study proposes an integrated feature representation framework EpiTEAmDNA based on the strategies of transfer learning and ensemble learning, which is evaluated on multiple DNA methylation types across 15 species. EpiTEAmDNA integrates convolutional neural network (CNN) and conventional machine learning methods, and shows improved performances than the existing DL-based methods on small datasets when no additional knowledge is available. The experimental data suggests that the EpiTEAmDNA models may be further improved via transfer learning based on additional knowledge. The evaluation experiments on the independent test datasets also suggest that the proposed EpiTEAmDNA framework outperforms the existing models in most prediction tasks of the 3 DNA methylation types across 15 species. The source code, pre-trained global model, and the EpiTEAmDNA feature representation framework are freely available at http://www.healthinformaticslab.org/supp/.

O-GlyThr: Prediction of human O-linked threonine glycosites using multi-feature fusion

O-linked glycosylation is one of the most complex post-translational modifications (PTM) of human proteins modulating various cellular metabolic and signaling pathways. Unlike N-glycosylation, the O-glycosylation has non-specific sequence features and unstable glycan core structure, which makes identification of O-glycosites more challenging either by experimental or computational methods. Biochemical experiments to identify O-glycosites in batches are technically and economically demanding. Therefore, development of computation-based methods is greatly warranted. This study constructed a prediction model based on feature fusion for O-glycosites linked to the threonine residues in Homo sapiens. In the training model, we collected and sorted out high-quality human protein data with O-linked threonine glycosites. Seven feature coding methods were fused to represent the sample sequence. By comparison of different algorithms, random forest was selected as the final classifier to construct the classification model. Through 5-fold cross-validation, the proposed model, namely O-GlyThr, performed satisfactorily on both training set (AUC: 0.9308) and independent validation dataset (AUC: 0.9323). Compared with previously published predictors, O-GlyThr achieved the highest ACC of 0.8475 on the independent test dataset. These results demonstrated the high competency of our predictor in identifying O-glycosites on threonine residues. Furthermore, a user-friendly webserver named O-GlyThr (http://cbcb.cdutcm.edu.cn/O-GlyThr/) was developed to assist glycobiologists in the research associated with glycosylation structure and function.

DeepAR: a novel deep learning-based hybrid framework for the interpretable prediction of androgen receptor antagonists

Drug resistance represents a major obstacle to therapeutic innovations and is a prevalent feature in prostate cancer (PCa). Androgen receptors (ARs) are the hallmark therapeutic target for prostate cancer modulation and AR antagonists have achieved great success. However, rapid emergence of resistance contributing to PCa progression is the ultimate burden of their long-term usage. Hence, the discovery and development of AR antagonists with capability to combat the resistance, remains an avenue for further exploration. Therefore, this study proposes a novel deep learning (DL)-based hybrid framework, named DeepAR, to accurately and rapidly identify AR antagonists by using only the SMILES notation. Specifically, DeepAR is capable of extracting and learning the key information embedded in AR antagonists. Firstly, we established a benchmark dataset by collecting active and inactive compounds against AR from the ChEMBL database. Based on this dataset, we developed and optimized a collection of baseline models by using a comprehensive set of well-known molecular descriptors and machine learning algorithms. Then, these baseline models were utilized for creating probabilistic features. Finally, these probabilistic features were combined and used for the construction of a meta-model based on a one-dimensional convolutional neural network. Experimental results indicated that DeepAR is a more accurate and stable approach for identifying AR antagonists in terms of the independent test dataset, by achieving an accuracy of 0.911 and MCC of 0.823. In addition, our proposed framework is able to provide feature importance information by leveraging a popular computational approach, named SHapley Additive exPlanations (SHAP). In the meanwhile, the characterization and analysis of potential AR antagonist candidates were achieved through the SHAP waterfall plot and molecular docking. The analysis inferred that N-heterocyclic moieties, halogenated substituents, and a cyano functional group were significant determinants of potential AR antagonists. Lastly, we implemented an online web server by using DeepAR (at http://pmlabstack.pythonanywhere.com/DeepAR). We anticipate that DeepAR could be a useful computational tool for community-wide facilitation of AR candidates from a large number of uncharacterized compounds.

Towards robustness and sensitivity of rapid Baijiu (Chinese liquor) discrimination using Raman spectroscopy and chemometrics: Dimension reduction, machine learning, and auxiliary sample

Rapid and non-invasive discrimination of two types of similar baijiu with the same alcohol content was investigated using chemometrics-assisted Raman spectroscopy. Baijiu samples manufactured across six months were subjected to Raman spectrum acquisition to form chronologically independent training and testing data sets for the optimization and evaluation of chemometric approaches. Discrimination capacities of machine learning (ML) algorithms were enhanced by the supervised spectral dimension reduction with linear discrimination analysis (LDA). The LDA-LightGBM model optimized with the binary training set led to an overall testing discrimination accuracy of 79%. Supplementing the Raman spectra of an auxiliary type of baijiu into the binary training data set considerably enhances the feature extraction capacity of LDA. Performances of the LDA-ML models were further elevated by the spectral preprocessing of “constant background subtraction” that shifts each spectrum downwards by its spectral intensity minimum. The chemometric approach that involved auxiliary training data, constant background subtraction, LDA dimension reduction, and ensemble learning classifier contributed to an overall testing discrimination accuracy of 97%. The proposed chemometric-assisted Raman spectroscopic approach facilitated robust and sensitive discrimination of highly similar baijiu products, which highlights the potential applications of the technique in manufacturing quality control and product authentication.

Matching single cells across modalities with contrastive learning and optimal transport.

Understanding the interactions between the biomolecules that govern cellular behaviors remains an emergent question in biology. Recent advances in single-cell technologies have enabled the simultaneous quantification of multiple biomolecules in the same cell, opening new avenues for understanding cellular complexity and heterogeneity. Still, the resulting multimodal single-cell datasets present unique challenges arising from the high dimensionality and multiple sources of acquisition noise. Computational methods able to match cells across different modalities offer an appealing alternative towards this goal. In this work, we propose MatchCLOT, a novel method for modality matching inspired by recent promising developments in contrastive learning and optimal transport. MatchCLOT uses contrastive learning to learn a common representation between two modalities and applies entropic optimal transport as an approximate maximum weight bipartite matching algorithm. Our model obtains state-of-the-art performance on two curated benchmarking datasets and an independent test dataset, improving the top scoring method by 26.1% while preserving the underlying biological structure of the multimodal data. Importantly, MatchCLOT offers high gains in computational time and memory that, in contrast to existing methods, allows it to scale well with the number of cells. As single-cell datasets become increasingly large, MatchCLOT offers an accurate and efficient solution to the problem of modality matching.

EdeepSADPr: an extensive deep-learning architecture for prediction of the in situ crosstalks of serine phosphorylation and ADP-ribosylation.

The in situ post-translational modification (PTM) crosstalk refers to the interactions between different types of PTMs that occur on the same residue site of a protein. The crosstalk sites generally have different characteristics from those with the single PTM type. Studies targeting the latter's features have been widely conducted, while studies on the former's characteristics are rare. For example, the characteristics of serine phosphorylation (pS) and serine ADP-ribosylation (SADPr) have been investigated, whereas those of their in situ crosstalks (pSADPr) are unknown. In this study, we collected 3,250 human pSADPr, 7,520 SADPr, 151,227pS and 80,096 unmodified serine sites and explored the features of the pSADPr sites. We found that the characteristics of pSADPr sites are more similar to those of SADPr compared to pS or unmodified serine sites. Moreover, the crosstalk sites are likely to be phosphorylated by some kinase families (e.g., AGC, CAMK, STE and TKL) rather than others (e.g., CK1 and CMGC). Additionally, we constructed three classifiers to predict pSADPr sites from the pS dataset, the SADPr dataset and the protein sequences separately. We built and evaluated five deep-learning classifiers in ten-fold cross-validation and independent test datasets. We also used the classifiers as base classifiers to develop a few stacking-based ensemble classifiers to improve performance. The best classifiers had the AUC values of 0.700, 0.914 and 0.954 for recognizing pSADPr sites from the SADPr, pS and unmodified serine sites, respectively. The lowest prediction accuracy was achieved by separating pSADPr and SADPr sites, which is consistent with the observation that pSADPr's characteristics are more similar to those of SADPr than the rest. Finally, we developed an online tool for extensively predicting human pSADPr sites based on the CNNOH classifier, dubbed EdeepSADPr. It is freely available through http://edeepsadpr.bioinfogo.org/. We expect our investigation will promote a comprehensive understanding of crosstalks.

Evaluating semi- and nonparametric regression algorithms in quantifying stem taper and volume with alternative test data selection strategies

Abstract Accurately quantifying stem taper is essential to predict diameter at any given height along the stem and to estimate tree volume for various sections of the stem. With increased computing power, semi- and nonparamatric methods have been proposed as alternative approaches for modelling tree taper. The main objective of this study was to assess the accuracy of stem taper predicted for four pine and four hardwood species by semi- and nonparametric models. Specifically, generalized additive models (GAM), random forests (RF) and regression-enhanced random forests (RERF) were compared with two widely-used parametric models: variable-exponent function (VAR) and segmented polynomial regression model (SEG). Test datasets selected from four different data splitting methods were used to examine the prediction accuracy of the models. Results showed that all examined models can be used to quantify stem taper and volume for all species when prediction is limited to be within the range of tree sizes used in model building. The nonparametric RF algorithm generally produced higher bias and lower precision than the semiparametric (GAM and RERF) and parametric models (VAR and SEG). For all species, GAM, VAR and SEG provided more robust predictions of stem taper than RF and RERF algorithms, especially when small or large trees were withheld for model testing. The data splitting strategies used in this work provide an efficient ‘stress test’ to evaluate model performance when collecting an independent test dataset is not feasible. The findings of this work provide additional insights for forest practitioners and resource managers to select appropriate methods in stem taper modelling.

An automatic immunofluorescence pattern classification framework for HEp-2 image based on supervised learning.

Immunofluorescence patterns of anti-nuclear antibodies (ANAs) on human epithelial cell (HEp-2) substrates are important biomarkers for the diagnosis of autoimmune diseases. There are growing clinical requirements for an automatic readout and classification of ANA immunofluorescence patterns for HEp-2 images following the taxonomy recommended by the International Consensus on Antinuclear Antibody Patterns (ICAP). In this study, a comprehensive collection of HEp-2 specimen images covering a broad range of ANA patterns was established and manually annotated by experienced laboratory experts. By utilizing a supervised learning methodology, an automatic immunofluorescence pattern classification framework for HEp-2 specimen images was developed. The framework consists of a module for HEp-2 cell detection and cell-level feature extraction, followed by an image-level classifier that is capable of recognizing all 14 classes of ANA immunofluorescence patterns as recommended by ICAP. Performance analysis indicated an accuracy of 92.05% on the validation dataset and 87% on an independent test dataset, which has surpassed the performance of human examiners on the same test dataset. The proposed framework is expected to contribute to the automatic ANA pattern recognition in clinical laboratories to facilitate efficient and precise diagnosis of autoimmune diseases.

Prediction of Proteins in Cerebrospinal Fluid and Application to Glioma Biomarker Identification.

Cerebrospinal fluid (CSF) proteins are very important because they can serve as biomarkers for central nervous system diseases. Although many CSF proteins have been identified with wet experiments, the identification of CSF proteins is still a challenge. In this paper, we propose a novel method to predict proteins in CSF based on protein features. A two-stage feature-selection method is employed to remove irrelevant features and redundant features. The deep neural network and bagging method are used to construct the model for the prediction of CSF proteins. The experiment results on the independent testing dataset demonstrate that our method performs better than other methods in the prediction of CSF proteins. Furthermore, our method is also applied to the identification of glioma biomarkers. A differentially expressed gene analysis is performed on the glioma data. After combining the analysis results with the prediction results of our model, the biomarkers of glioma are identified successfully.

Spatial reconstruction of long-term (2003–2020) sea surface pCO2 in the South China Sea using a machine-learning-based regression method aided by empirical orthogonal function analysis

Abstract. The South China Sea (SCS) is the largest marginal sea of the North Pacific Ocean, where intensive field observations, including mappings of the sea surface partial pressure of CO2 (pCO2), have been conducted over the last 2 decades. It is one of the most studied marginal seas in terms of carbon cycling and could thus be a model system for marginal sea carbon research. However, the cruise-based sea surface pCO2 datasets are still temporally and spatially sparse. Using a machine-learning-based method facilitated by empirical orthogonal function (EOF) analysis, this study provides a reconstructed dataset of the monthly sea surface pCO2 in the SCS with a reasonably high spatial resolution (0.05∘ × 0.05∘) and temporal coverage between 2003 and 2020. The data input to our model includes remote-sensing-derived sea surface salinity, sea surface temperature, and chlorophyll, the spatial pattern of pCO2 constrained by EOF, atmospheric pCO2, and time labels (month). We validated our reconstruction with three independent testing datasets that are not involved in the model training. Among them, Test 1 includes 10 % of our in situ data, Test 2 contains four independent in situ datasets corresponding to the four seasons, and Test 3 is an in situ monthly dataset available from 2003–2019 at the South East Asia Time-series Study (SEATs) station located in the northern basin of the SCS. Our Test 1 validation demonstrated that the reconstructed pCO2 field successfully simulated the spatial and temporal patterns of sea surface pCO2 observations. The root mean square error (RMSE) between our reconstructed data and in situ data in Test 1 averaged ∼10 µatm, which is much smaller (by ∼50 %) than that between the remote-sensing-derived data and in situ data. Test 2 verified the accuracy of our retrieval algorithm in months lacking observations, showing a relatively small bias (RMSE of ∼8 µatm). Test 3 evaluated the accuracy of the reconstructed long-term trend, showing that, at the SEATs station, the difference between the reconstructed pCO2 and in situ data ranged from −10 to 4 µatm (−2.5 % to 1 %). In addition to the typical machine learning performance metrics, we assessed the uncertainty resulting from reconstruction bias and its feature sensitivity. These validations and uncertainty analyses strongly suggest that our reconstruction effectively captures the main spatial and temporal features of sea surface pCO2 distributions in the SCS. Using the reconstructed dataset, we show the long-term trends of sea surface pCO2 in five subregions of the SCS with differing physicobiogeochemical characteristics. We show that mesoscale processes such as the Pearl River plume and China coastal currents significantly impact sea surface pCO2 in the SCS during different seasons. While the SCS is overall a weak source of atmospheric CO2, the northern SCS acts as a sink, showing a trend of increasing strength over the past 2 decades. The data used in this article are available at https://doi.org/10.57760/sciencedb.02050 (Wang and Dai, 2022).

Segmentation of thyroid glands and nodules in ultrasound images using the improved U-Net architecture

BackgroundIdentifying thyroid nodules’ boundaries is crucial for making an accurate clinical assessment. However, manual segmentation is time-consuming. This paper utilized U-Net and its improved methods to automatically segment thyroid nodules and glands.MethodsThe 5822 ultrasound images used in the experiment came from two centers, 4658 images were used as the training dataset, and 1164 images were used as the independent mixed test dataset finally. Based on U-Net, deformable-pyramid split-attention residual U-Net (DSRU-Net) by introducing ResNeSt block, atrous spatial pyramid pooling, and deformable convolution v3 was proposed. This method combined context information and extracts features of interest better, and had advantages in segmenting nodules and glands of different shapes and sizes.ResultsDSRU-Net obtained 85.8% mean Intersection over Union, 92.5% mean dice coefficient and 94.1% nodule dice coefficient, which were increased by 1.8%, 1.3% and 1.9% compared with U-Net.ConclusionsOur method is more capable of identifying and segmenting glands and nodules than the original method, as shown by the results of correlational studies.

Prediction of Deep Myometrial Infiltration, Clinical Risk Category, Histological Type, and Lymphovascular Space Invasion in Women with Endometrial Cancer Based on Clinical and T2-Weighted MRI Radiomic Features.

To predict deep myometrial infiltration (DMI), clinical risk category, histological type, and lymphovascular space invasion (LVSI) in women with endometrial cancer using machine learning classification methods based on clinical and image signatures from T2-weighted MR images. A training dataset containing 413 patients and an independent testing dataset consisting of 82 cases were employed in this retrospective study. Manual segmentation of the whole tumor volume on sagittal T2-weighted MRI was performed. Clinical and radiomic features were extracted to predict: (i) DMI of endometrial cancer patients, (ii) endometrial cancer clinical high-risk level, (iii) histological subtype of tumor, and (iv) presence of LVSI. A classification model with different automatically selected hyperparameter values was created. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve, F1 score, average recall, and average precision were calculated to evaluate different models. Based on the independent external testing dataset, the AUCs for DMI, high-risk endometrial cancer, endometrial histological type, and LVSI classification were 0.79, 0.82, 0.91, and 0.85, respectively. The corresponding 95% confidence intervals (CI) of the AUCs were [0.69, 0.89], [0.75, 0.91], [0.83, 0.97], and [0.77, 0.93], respectively. It is possible to classify endometrial cancer DMI, risk, histology type, and LVSI using different machine learning methods.

ASLncR: a novel computational tool for prediction of abiotic stress-responsive long non-coding RNAs in plants.

Abiotic stresses are detrimental to plant growth and development and have a major negative impact on crop yields. A growing body of evidence indicates that a large number of long non-coding RNAs (lncRNAs) are key to many abiotic stress responses. Thus, identifying abiotic stress-responsive lncRNAs is essential in crop breeding programs in order to develop cropcultivars resistant to abiotic stresses. In this study, we have developed the first machine learning-based computational model for predicting abiotic stress-responsive lncRNAs. The lncRNA sequences which were responsive and non-responsive to abiotic stresses served as the two classes of the dataset for binary classification using the machine learning algorithms. The training dataset was created using 263 stress-responsive and 263 non-stress-responsive sequences, whereas the independent test set consists of 101 sequences from both classes. As the machine learning model can adopt only the numeric data, the Kmer features ranging from sizes 1 to 6 were utilized to represent lncRNAs in numeric form. To select important features, four different feature selection strategies were utilized. Among the seven learning algorithms, the support vector machine (SVM) achieved the highest cross-validation accuracy with the selected feature sets. The observed 5-fold cross-validation accuracy, AU-ROC, and AU-PRC were found to be 68.84, 72.78, and 75.86%, respectively. Furthermore, the robustness of the developed model (SVM with the selected feature) was evaluated using an independent test dataset, where the overall accuracy, AU-ROC, and AU-PRC were found to be 76.23, 87.71, and 88.49%, respectively. The developed computational approach was also implemented in an online prediction tool ASLncR accessible at https://iasri-sg.icar.gov.in/aslncr/ . The proposed computational model and the developed prediction tool are believed to supplement the existing effort for the identification of abiotic stress-responsive lncRNAs in plants.

Improving protein-protein interaction site prediction using deep residual neural network

Accurate identification of protein-protein interaction (PPI) sites is significantly important for understanding the mechanism of life and developing new drugs. However, it is expensive and time-consuming to identify PPI sites using wet-lab experiments. Developing computational methods is a new road to identify PPI sites, which can accelerate the procedure of PPI-related research. In this study, we propose a novel deep learning-based method (called D-PPIsite) to improve the accuracy of sequence-based PPI site prediction. In D-PPIsite, four discriminative sequence-driven features, i.e., position specific scoring matrix, relative solvent accessibility, position information and physical properties, are employed to feed into a well-designed deep learning module, consisting of convolutional, squeeze and excitation, and fully connected layers, to learn a prediction model. To reduce the risk of a single prediction model getting stuck in local optima, multiple prediction models with different initialization parameters are selected and integrated into one final model using the mean ensemble strategy. Experimental results on five independent testing data sets demonstrate that the proposed D-PPIsite can achieve an average accuracy of 80.2% and precision of 36.9%, covering 53.5% of all PPI sites while achieving the average Matthews correlation coefficient value (0.330) that is significantly higher than most of existing state-of-the-art prediction methods. We implement a new standalone-version predictor for predicting PPI sites, which is freely available at https://github.com/MingDongup/D-PPIsite for academic use.

Deep Learning With an Attention Mechanism for Differentiating the Origin of Brain Metastasis Using MR images.

Brain metastasis (BM) is a serious neurological complication of cancer of different origins. The value of deep learning (DL) to identify multiple types of primary origins remains unclear. To distinguish primary site of BM and identify the best DL models. Retrospective. A total of 449 BM derived from 214 patients (49.5% for female, mean age 58 years) (100 from small cell lung cancer [SCLC], 125 from non-small cell lung cancer [NSCLC], 116 from breast cancer [BC], and 108 from gastrointestinal cancer [GIC]) were included. A 3-T, T1 turbo spin echo (T1-TSE), T2-TSE, T2FLAIR-TSE, DWI echo-planar imaging (DWI-EPI) and contrast-enhanced T1-TSE (CE T1-TSE). Lesions were divided into training (n=285, 153 patients), testing (n=122, 93 patients), and independent testing cohorts (n=42, 34 patients). Three-dimensional residual network (3D-ResNet), named 3D ResNet6 and 3D ResNet 18, was proposed for identifying the four origins based on single MRI and combined MRI (T1WI + T2-FLAIR + DWI, CE-T1WI + DWI, CE-T1WI + T2WI + DWI). DL model was used to distinguish lung cancer from non-lung cancer; then SCLC vs. NSCLC for lung cancer classification and BC vs. GIC for non-lung cancer classification was performed. A subjective visual analysis was implemented and compared with DL models. Gradient-weighted class activation mapping (Grad-CAM) was used to visualize the model by heatmaps. The area under the receiver operating characteristics curve (AUC) assess each classification performance. 3D ResNet18 with Grad-CAM and AIC showed better performance than 3DResNet6, 3DResNet18 and the radiologist for distinguishing lung cancer from non-lung cancer, SCLC from NSCLC, and BC from GIC. For single MRI sequence, T1WI, DWI, and CE-T1WI performed best for lung cancer vs. non-lung cancer, SCLC vs. NSCLC, and BC vs. GIC classifications. The AUC ranged from 0.675 to 0.876 and from 0.684 to 0.800 regarding the testing and independent testing datasets, respectively. For combined MRI sequences, the combination of CE-T1WI + T2WI + DWI performed better for BC vs. GIC (AUCs of 0.788 and 0.848 on testing and independent testing datasets, respectively), while the combined MRI approach (T1WI + T2-FLAIR + DWI, CE-T1WI + DWI) could not achieve higher AUCs for lung cancer vs. non-lung cancer, SCLC vs. NSCLC. Grad-CAM helped for model visualization by heatmaps that focused on tumor regions. DL models may help to distinguish the origins of BM based on MRI data. 3 TECHNICAL EFFICACY: Stage 2.

Deep learning based deep-sea automatic image enhancement and animal species classification

The automatic classification of marine species based on images is a challenging task for which multiple solutions have been increasingly provided in the past two decades. Oceans are complex ecosystems, difficult to access, and often the images obtained are of low quality. In such cases, animal classification becomes tedious. Therefore, it is often necessary to apply enhancement or pre-processing techniques to the images, before applying classification algorithms. In this work, we propose an image enhancement and classification pipeline that allows automated processing of images from benthic moving platforms. Deep-sea (870 m depth) fauna was targeted in footage taken by the crawler “Wally” (an Internet Operated Vehicle), within the Ocean Network Canada (ONC) area of Barkley Canyon (Vancouver, BC; Canada). The image enhancement process consists mainly of a convolutional residual network, capable of generating enhanced images from a set of raw images. The images generated by the trained convolutional residual network obtained high values in metrics for underwater imagery assessment such as UIQM (~ 2.585) and UCIQE (2.406). The highest SSIM and PSNR values were also obtained when compared to the original dataset. The entire process has shown good classification results on an independent test data set, with an accuracy value of 66.44% and an Area Under the ROC Curve (AUROC) value of 82.91%, which were subsequently improved to 79.44% and 88.64% for accuracy and AUROC respectively. These results obtained with the enhanced images are quite promising and superior to those obtained with the non-enhanced datasets, paving the strategy for the on-board real-time processing of crawler imaging, and outperforming those published in previous papers.

ECAmyloid: An amyloid predictor based on ensemble learning and comprehensive sequence-derived features

Amyloid fibrils formed by the mis-aggregation of amyloid proteins can lead to neuronal degenerations in the Alzheimer's disease. Predicting amyloid proteins not only contributes to understanding physicochemical properties and formation mechanism of amyloid proteins, but also has significant implications in the amyloid disease treatment and the development of a new purpose for amyloid materials. In this study, an ensemble learning model with sequence-derived features, ECAmyloid, is proposed to identify amyloids. The sequence-derived features including Pseudo Position Specificity Score Matrix (Pse-PSSM), Split Amino Acid Composition (SAAC), Solvent Accessibility (SA), and Secondary Structure Information (SSI) are employed to incorporate sequence composition, evolutionary and structural information. The individual learners of the ensemble learning model are selected by an increment classifier selection strategy. The final prediction results are determined by voting of prediction results of multiple individual learners. In view of the imbalanced benchmark dataset, the Synthetic Minority Over-sampling Technique (SMOTE) is adopted to generate positive samples. To eliminate irrelevant features and redundant features, correlation-based feature subset (CFS) selection combined with a heuristic search strategy is performed to obtain the optimal feature subset. Experimental results indicate that the ensemble classifier achieves an accuracy of 98.29%, a sensitivity of 0.992, a specificity of 0.974 on the training dataset using the 10-fold cross validation, far higher than the results obtained by its individual learners. Compared with the original feature set, the accuracy, sensitivity, specificity, MCC, F1-score, G-Mean of the ensemble method trained by the optimal feature subset are improved by 1.05%, 0.012, 0.01, 0.021, 0.011 and 0.011, respectively. Moreover, the comparison results with existing methods on two same independent test datasets demonstrate that the proposed method is an effective and promising predictor for large-scale determination of amyloid proteins. The data and code used to develop ECAmyloid has been shared to Github, and can be freely downloaded at https://github.com/KOALA-L/ECAmyloid.git.

Dual center validation of deep learning for automated multi-label segmentation of thoracic anatomy in bedside chest radiographs

Background and ObjectivesBedside chest radiographs (CXRs) are challenging to interpret but important for monitoring cardiothoracic disease and invasive therapy devices in critical care and emergency medicine. Taking surrounding anatomy into account is likely to improve the diagnostic accuracy of artificial intelligence and bring its performance closer to that of a radiologist. Therefore, we aimed to develop a deep convolutional neural network for efficient automatic anatomy segmentation of bedside CXRs. MethodsTo improve the efficiency of the segmentation process, we introduced a "human-in-the-loop" segmentation workflow with an active learning approach, looking at five major anatomical structures in the chest (heart, lungs, mediastinum, trachea, and clavicles). This allowed us to decrease the time needed for segmentation by 32% and select the most complex cases to utilize human expert annotators efficiently. After annotation of 2,000 CXRs from different Level 1 medical centers at Charité – University Hospital Berlin, there was no relevant improvement in model performance, and the annotation process was stopped. A 5-layer U-ResNet was trained for 150 epochs using a combined soft Dice similarity coefficient (DSC) and cross-entropy as a loss function. DSC, Jaccard index (JI), Hausdorff distance (HD) in mm, and average symmetric surface distance (ASSD) in mm were used to assess model performance. External validation was performed using an independent external test dataset from Aachen University Hospital (n = 20). ResultsThe final training, validation, and testing dataset consisted of 1900/50/50 segmentation masks for each anatomical structure. Our model achieved a mean DSC/JI/HD/ASSD of 0.93/0.88/32.1/5.8 for the lung, 0.92/0.86/21.65/4.85 for the mediastinum, 0.91/0.84/11.83/1.35 for the clavicles, 0.9/0.85/9.6/2.19 for the trachea, and 0.88/0.8/31.74/8.73 for the heart. Validation using the external dataset showed an overall robust performance of our algorithm. ConclusionsUsing an efficient computer-aided segmentation method with active learning, our anatomy-based model achieves comparable performance to state-of-the-art approaches. Instead of only segmenting the non-overlapping portions of the organs, as previous studies did, a closer approximation to actual anatomy is achieved by segmenting along the natural anatomical borders. This novel anatomy approach could be useful for developing pathology models for accurate and quantifiable diagnosis.