SemOpenAlex |

SemOpenAlex

Matches in SemOpenAlex for { <https://semopenalex.org/work/W4311895184> ?p ?o ?g. }

Showing items 1 to 45 of 45 with 100 items per page.

W4311895184 abstract "Article Figures and data Abstract Editor's evaluation Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Mice are the most commonly used model animals for itch research and for development of anti-itch drugs. Most laboratories manually quantify mouse scratching behavior to assess itch intensity. This process is labor-intensive and limits large-scale genetic or drug screenings. In this study, we developed a new system, Scratch-AID (Automatic Itch Detection), which could automatically identify and quantify mouse scratching behavior with high accuracy. Our system included a custom-designed videotaping box to ensure high-quality and replicable mouse behavior recording and a convolutional recurrent neural network trained with frame-labeled mouse scratching behavior videos, induced by nape injection of chloroquine. The best trained network achieved 97.6% recall and 96.9% precision on previously unseen test videos. Remarkably, Scratch-AID could reliably identify scratching behavior in other major mouse itch models, including the acute cheek model, the histaminergic model, and a chronic itch model. Moreover, our system detected significant differences in scratching behavior between control and mice treated with an anti-itch drug. Taken together, we have established a novel deep learning-based system that could replace manual quantification for mouse scratching behavior in different itch models and for drug screening. Editor's evaluation Scratch assays are the gold standard for measuring itch in rodents. However, the current limitation is that this is performed manually which is enormously taxing in terms of hours spent counting scratching bouts. The authors have developed a valuable automatic system to quantify scratch behavior with high accuracy and provided a valuable tool for the field. This will be resourceful for the greater itch biology community. https://doi.org/10.7554/eLife.84042.sa0 Decision letter eLife's review process Introduction Itch is a disturbing symptom associated with skin diseases, immune problems, systemic diseases, and mental disorders (Cevikbas and Lerner, 2020; Hong et al., 2011; Kremer et al., 2020; Ständer et al., 2007). Chronic itch affects about 13–17% of the population (Matterne et al., 2009; Weisshaar and Dalgard, 2009), severely worsening the quality of life of affected patients. Unfortunately, treatment options for many chronic itch conditions are still limited (Yosipovitch et al., 2018; Yu et al., 2021). Mice are the most widely used model animals for studying itch mechanisms and for developing new preclinical anti-itch drugs (Han et al., 2013; Liu et al., 2009; Solinski et al., 2019; Sun and Chen, 2007). Since itch is an unpleasant sensation that provokes the desire to scratch (Ikoma et al., 2006), scratching behavior has been assessed as a proxy for itch intensity in mice (Liu et al., 2009; Morita et al., 2015). Till now, this quantification process has been mainly conducted by watching videos and manually counting scratching bouts or the total scratching time, which is tedious and time consuming, unavoidably introduces human errors and bias, and limits the large-scale genetic or drug screenings. Given the biological importance and the obvious need, several research groups have tried different strategies to automate this process, including an acoustic recording method (Elliott et al., 2017), a method using magnetic field and metal ring to detect paw movement (Mu et al., 2017), and several video analysis-based approaches (Bohnslav et al., 2021; Kobayashi et al., 2021; Park et al., 2019; Sakamoto et al., 2022). Nevertheless, these methods have not been widely adopted by other research laboratories, due to the uncertain performance of the trained models in different lab environments, the requirement of specialized equipment, and/or inadequate evaluation of these methods in different mouse itch models. In recent years, with the rapid progress in the field of artificial intelligence, deep learning has been applied in various scientific research areas. For example, convolutional neural networks (CNN) are widely used in computer visual recognition tasks (Gu et al., 2018), whereas recurrent neural networks (RNN) (Graves, 2013) are developed for analyzing temporal dynamic features. Moreover, rapid improvement of computing power, especially in the graphics processing unit capacity, together with new open-source deep learning libraries, such as PyTorch (Paszke et al., 2019), Keras (Gulli and Pal, 2017), and Tensorflow (Abadi et al., 2016), have greatly accelerated broad applications of deep learning. Animal behavior analysis is one of the research areas benefiting from the applications of deep learning. For example, DeepLabCut can track different body parts in freely moving animals for behavior analysis (Mathis et al., 2018). DeepEthogram recognizes and annotates different behaviors of mice and flies (Bohnslav et al., 2021). These examples support the proof-of-principle that deep-learning is a powerful avenue for automating animal behavior analysis. Nevertheless, for a given animal behavior, like mouse scratching, a designated method, which achieves high sensitivity, specificity, and generalization to replace human observers, still needs to be established. To meet this challenge, we developed a new deep learning-based system, Scratch-AID (Automatic Itch Detection), which achieved automatic quantification of mouse scratching behavior with high accuracy. We first designed a videotaping box to acquire high-quality recording of mouse behavior in a reproducible environment from the bottom. We recorded 40 videos of 10 adult wildtype mice (5 males and 5 females) after nape injection of a non-histamine pruritogen, chloroquine (CQ), and manually labeled all video frames as the reference. We then designed a convolutional recurrent neural network (CRNN) by combining CNN and RNN and trained it with 32 scratching videos from 8 randomly picked mice. We obtained a series of prediction models using different training parameters and evaluated these models with test videos (eight unseen test videos from the two remaining mice). The best trained model achieved 97.6% recall and 96.9% precision on test videos, similar to the manual quantification results. Impressively, Scratch-AID could also quantify scratching behavior from other major acute and chronic itch models with high accuracy. Lastly, we applied Scratch-AID in an anti-itch drug screening paradigm and found that it reliably detected the drug effect. In summary, we have established a new system for accurate automatic quantification of mouse scratching behavior. Based on the performances, Scratch-AID could replace manual quantification for various mouse itch models and for drug or genetic screenings. Results The overall workflow Our workflow to develop a new system for detection and quantification of mouse scratching behavior consists of four major steps (Figure 1A): (1) Videotape mouse scratching behavior induced by an acute nape itch model; (2) Manually annotate scratching frames in all videos for training and test datasets; (3) Design a deep learning neural network; train this network with randomly selected training videos and adjust different training parameters; and evaluate the performance of the trained neural networks on test videos; (4) Evaluate the generalization of the trained neural network in additional itch models and a drug screening paradigm. Figure 1 with 1 supplement see all Download asset Open asset The overall workflow and building a customized videotaping box for mouse scratching behavior recording. (A) A diagram showing the workflow to develop a deep learning-based system for automatic detection and quantification of mouse scratching behavior. (B) An image of the designed videotaping box for high-quality video recording of mouse scratching behavior. Scale bar, 5 cm. (C) A cartoon showing the acute itch model induced by the chloroquine (CQ) injection in the nape, followed by video recording in the customized videotaping box. (D) Representative images showing different phases (P1–P4) of a scratching train (upper). Red arrows indicate the scratching hind paw. A cartoon showing the dynamic movement of the scratching hind paw in a scratching train (bottom). The cycle of scratching bout (P2) and paw licking (P3) may repeat once or more times in a scratching train. Scale bar, 1 cm. (E) The total number of scratching trains in each video. (F) The distribution of scratching train duration (n = 1135 scratching trains). The inset is the zoom-in of the red rectangle. Design a videotaping box for high-quality and reproducible recording of mouse scratching behavior High-quality videos recorded from a reproducible environment are critical for stable performance of trained prediction models and for adoption by other research laboratories. Thus, we designed a mouse videotaping box for such a purpose. It consisted of two boxes with white acrylic walls joined by a transparent acrylic floor (Figure 1B). The top (‘mouse’) box (Length × Width × Height = 14.68 × 14.68 × 5 cm) had a lid for mouse access. The bottom (‘camera’) box (Length × Width × Height = 14.68 × 14.68 × 23.6 cm) had a door for access to the camera (Logitech C920e Business Webcam). The walls and the lid of the box were non-transparent to minimize interference from outside visual stimuli. Ambient light penetrated the walls and provided sufficient illumination for behavior recording. A mouse could freely move inside the top box, and a camera recorded mouse behavior from the bottom (30 frames/s). Compared to the top or side views, the bottom view can clearly capture the key body parts involved in scratching behavior, such as the scratching hind paw and mouth, and their movements in great details (Video 1). The magnification, resolution, and brightness of the video can be adjusted by the camera recording software (Logitech C920e Business Webcam driver and software) to achieve consistent video recording. In short, this customized videotaping box allows high-quality video recording of mouse scratching and other behaviors in a stable and reproducible environment. Video 1 Download asset This video cannot be played in place because your browser does support HTML5 video. You may still download the video for offline viewing. Download as MPEG-4 Download as WebM Download as Ogg An example of a mouse scratching train recorded by the designed videotaping box. Spontaneous scratching is a rare event under normal conditions in mice. Itch sensation and scratching are usually induced by different itch models for research. Common mouse itch models are classified as cheek or nape, histaminergic or non-histaminergic, and acute or chronic, based on the body location where itch sensation is evoked, the kind of pruritogens, and the duration of itch sensation (Ikoma et al., 2006; Liu and Dong, 2015; Shimada and LaMotte, 2008; Thurmond et al., 2008). We first used an acute nape itch model induced by a non-histaminergic pruritogen, CQ, because it triggered immediate and robust scratching behavior in mice (Liu et al., 2009). After intradermal injection of CQ (200 μg in 15 μl saline) into left or right nape of the mice, a 20-min video was recorded using the customized videotaping box (Figure 1C). In response to CQ injection, mice scratched the affected skin region using their ipsilateral hind paw. Evoked mice displayed multiple episodes of scratching, which were separated by non-scratching phases. Each episode of scratching, here defined as a scratching train, usually contained four phases: start (lifting the scratching hind paw toward the affected skin), scratching bout (rhythmic movement of hind paw against the affected skin), paw licking (putting the scratching hind paw into the mouth and licking), and end (putting down the scratching hind paw back to the floor) (Figure 1D and Video 1). The cycle of scratching bout and paw licking might occur once or repeat multiple times, depending on the itch intensity and the internal state of the mouse. The time from the start to the end of a given scratching train is defined as the duration. The total scratching time is the sum of durations of all scratching trains, which is an effective parameter to quantify the scratching behavior and assess itch intensity. Video annotation Forty scratching videos from 10 adult wildtype C57 mice (5 males and 5 females, 2- to 3-month-old) were recorded (Supplementary file 1). For neural network training and testing and for comparing performances between trained neural networks and manual quantification, two methods were used to annotate mouse scratching behavior in these videos. The first method (manual annotation) was to watch these videos at normal (1×) speed (30 frames/s) and label the start and end time points (converted into frame numbers for subsequent analysis) of each scratching train. This is consistent with the field common practice for manually quantifying mouse scratching behavior. Our manual annotation results were produced by 10 human observers, thus reflecting an averaged precision of the manual quantification process. The second method (reference annotation, or ground-truth annotation) was to accurately determine the start and end of each scratching train by analyzing each video frame-by-frame. The reference annotation of the 40 videos were used as the training and test datasets. The total number of scratching train in each video and the distribution of scratching train durations were quantified (Figure 1E, F and Supplementary file 2). Deep learning neural network design and model training Mouse scratching behavior displayed unique dynamic (temporal) and static (spatial) features, which were highlighted by tracking the key body parts using DeepLabCut (Mathis et al., 2018; Figure 1—figure supplement 1; Video 2). One of the most obvious dynamic features was the rhythmic movement of the scratching hind paw (Figure 1—figure supplement 1A, B). Some unique static features included the relative positional relationships between the scratching hind paw and other body parts (Figure 1—figure supplement 1C–F). To fully capture these dynamic and static features, we designed a CRNN to take advantage of the different strengths of CNN and RNN (Figure 2A and Figure 2—figure supplement 1A). The CRNN contained a CNN (ResNet-18 He et al., 2016; Figure 2—figure supplement 1B) that extracts static features, such as the relative position of different body parts, an RNN (two-layer bidirectional gated recurrent unit [GRU]; Dey and Salem, 2017; Figure 2—figure supplement 1C) that extracts dynamic features, such as the rhythmic movement of the scratching hind paw in consecutive frames, and a fully connected layer (the classifier) to combine the features extracted by both CNN and RNN and generate the prediction output. Video 2 Download asset This video cannot be played in place because your browser does support HTML5 video. You may still download the video for offline viewing. Download as MPEG-4 Download as WebM Download as Ogg Key body parts tracked by DeepLabCut showing the dynamic and static features of mouse scratching behavior. Figure 2 with 2 supplements see all Download asset Open asset Deep learning neural network design and training. (A) Cartoon showing the architecture of designed deep learning neural network consisting of the combination of convolutional neural networks (CNN), recurrent neural networks (RNN), and classifier. (B) Cartoon showing the preparation of inputs for the training dataset. Consecutive N frames were selected as one input for training. The interval between two adjacent inputs in a video was 4–10 frames. (C) The information of a sample training and test datasets. The training loss decreased (D) while the accuracy increased (E) during the training process with different input length (N = 3, 5, 7, 13, 23, 45 frames). The inset is the zoom-in of part of the figure. The 40 videos were randomly split into two parts, 80% of them (32 videos from 8 mice) were assigned to the training dataset and 20% of them (8 videos from 2 mice) to the test dataset (Figure 2C). Each video was converted into individual frames, and each frame was classified as ‘scratching’ (within a scratching train) or ‘non-scratching’ (out of a scratching train) based on the reference annotation. For the training dataset preparation, N consecutive frames (a parameter adjusted for optimal model performance) were selected as one input to capture the dynamic features of scratching (Figure 2B). To avoid large sets of redundancy in the training dataset, two adjacent inputs were apart between 4 and 10 frames (Figure 2B). An input was labeled as ‘scratching’ (class 1) if more than half of frames (N/2) in the input were scratching frames; otherwise labeled as ‘non-scratching’ (class 0) (Figure 2B). Since the dynamic features of scratching behavior spanned multiple frames, the input length (N frames) would be a critical training parameter. In CQ triggered acute nape itch model, the average duration of one cycle of scratching bout and paw licking was around 30 frames (Figure 2—figure supplement 2A, B). Thus, we tested a range of input length from 3 to 45 frames for model training. During the training process, the loss (discrepancy between model prediction and reference annotation) decreased, and the prediction accuracy (correct prediction of both scratching and non-scratching frames/all frames) increased (Figure 2D, E). After 10 epochs (one epoch means the training covers the complete training dataset for one round), the accuracy reached a plateau (Figure 2E). The prediction accuracies were more than 0.98 for all input length, improving slightly with the increase of the input length (Figure 2E). These results demonstrate that the designed CRNN network works very efficiently to capture the scratching features and recognize scratching behavior in the training dataset. Model evaluation on test datasets We evaluated the performance of the trained prediction models on eight unseen test videos. First, similar to what described above, each test video was converted into inputs with ‘N’ frames (the same ‘N’ was used for training and test), except that the two adjacent inputs were only 1 frame apart. Second, the trained neural network predicted each input to be ‘scratching’ or ‘non-scratching’. Third, to convert the prediction from one input (containing N frames) into the prediction for each individual frame, we used the following rule: the prediction of the middle frame of each input would be the same as that of the input. For example, if an input was predicted as ‘scratching’, then the middle frame of this input would be a ‘scratching’ frame. This conversion predicted each frame of tested videos as ‘scratching’ or ‘non-scratching’ expect for the few frames at the beginning or at the end of a video (see method for the missing data interpretation). Fourth, recall (the number of correctly predicted scratching frames/the number of reference scratching frames), precision (the number of correctly predicted scratching frames/the number of all predicted scratching frames), and F1 score (2*recall*precision/(recall + precision)) were calculated. Compared to the overall accuracy (correct prediction of both scratching and non-scratching frames/all frames), the recall, precision, and F1 score give a more precise and in depth evaluation of a model’s performance (Powers, 2020), especially when scratching is a relative rare event in a video. To rule out the possibility that the good performance of our models was due to a specific combination of the training and test datasets, we rotated the training and test videos for cross-validation. For all different combination of training and test videos, F1 scores were all above 0.9 (Figure 3—figure supplement 1A, B), supporting the stable and high performance of our prediction models. In addition, the prediction model performed better with the increased input length (Figure 3—figure supplement 1C). The best model, the one trained with videos 1–32 with the input length of 45 (N = 45), was selected for additional analyses and tests. The average recall and precision of the best model were 97.6% and 96.9%, respectively, for the eight test videos (Figure 3A), and the recall and precision for individual videos were above 95% in most cases (Figure 3B). This performance was similar to or even slightly better than that of manual annotation, which had an average recall and precision of 95.1% and 94.2%, respectively (Figure 3C, D). When comparing the total scratching time to the reference annotation, the average discrepancy of the model prediction was 1.9% whereas that of manual annotation was 2.1% (Figure 3E). The correlation between the model prediction and the reference annotation was 0.98, similar to the manual annotation results (Figure 3F). When examining the probability traces of model prediction and the reference annotation (one example shown in Figure 3G), we found that the model successfully recognized almost all the scratching trains in test videos, and that the prediction of the start and end of each scratching train aligned well with the reference annotation (Figure 3H). Taken together, these results demonstrate the high reliability and accuracy of our model to recognize and quantify mouse scratching behavior of new videos. Figure 3 with 5 supplements see all Download asset Open asset Performance of the best model on test videos. The recall, precision, and F1 score of the best model on average (A) or in individual videos (B). The recall, precision, and F1 score of manual annotation on average (C) or in individual videos (D). (E) The comparison among model prediction, manual quantification, and the reference annotation. The reference annotation is normalized to 100% shown as the red line. (F) The correlations between model prediction or manual quantification and reference annotation. R2, Pearson correlation coefficient. (G) An example scratching probability trace (red curve) predicted by the model and aligned with the reference annotation (green bar). (H) The two zoom-ins from (G) showing the nice alignment between the model prediction and the reference annotation. The trained neural network model focused on the scratching hind paw to recognize mouse scratching behavior How did the trained neural network model recognize mouse scratching behavior and distinguish them from other behaviors? Although deep learning neural networks are processed as a black box, saliency maps can give some hints (Selvaraju et al., 2017), because they plot which parts of each frame (pixels) were mainly used during model prediction. The most salient parts were centered around the scratching hind paw in the scratching frames (Figure 4A, B and Figure 4—figure supplement 1A), suggesting that the prediction model focused on the features of the scratching hind paw. In some scratching frames, other body parts were also highlighted, such as the two front paws (Figure 4B), suggesting the model also utilized the positional relationship of these body parts to recognize scratching behavior. In contrast, for other ‘non-scratching’ mouse behaviors, such as wiping, grooming, rearing, and locomotion, the ‘salient’ parts showed no clear association with particular mouse body parts (Figure 4C–F and Figure 4—figure supplement 1B–E). Together, these saliency maps indicate that the trained neural network learns to focus on the dynamic and static features of scratching for the prediction of mouse scratching behavior. Figure 4 with 1 supplement see all Download asset Open asset The prediction model focused on the scratching hind paw for scratching behavior recognition. (A, B) Saliency map showing the gradient value of each pixel of scratching frames during mouse scratching behavior prediction by the best model. The model focused on the scratching hind paw (A, B) and other body parts, such as front paws (B).Scale bar, 2 cm. Saliency map showing the gradient value of each pixel of wiping (C), grooming (D), rearing (E), and locomotion (F) frames during mouse scratching behavior prediction by the model. Model prediction error analysis To further understand the performance of the best trained neural network model, we systematically analyzed its prediction errors in eight test videos (Figure 3—figure supplement 2) and compared to those from the manual quantification (Figure 3—figure supplement 3) and other trained models. We classified the prediction errors into five categories: Type 1, false positive (non-scratching region was predicted as a scratching train); Type 2, false negative (a real scratching train was not recognized); Type 3, blurred boundary (the prediction of the start or end of a scratching train was shifted); Type 4, missed interval (two or more adjacent scratching trains were predicted as one scratching train); and Type 5 split scratching train (one scratching train was predicted as two or more scratching trains) (Figure 3—figure supplement 2A). We found that the dominant prediction error of the trained neural network model was Type 3 error, accounting for around 3% of the total scratching frames, followed by Type 2 and 5 errors accounting for around 1% (Figure 3—figure supplement 2B). For manual quantification, the major errors came from Types 3 and 4, accounting for 10% and 8% of the total scratching frames, respectively (Figure 3—figure supplement 3A). For Type 1 error of the model prediction (Figure 3—figure supplement 2C1–C3), the durations of all false positive scratching trains were shorter than 10 frames (0.3 s) (Figure 3—figure supplement 2C2) and temporarily close to a real scratching train (within 30 frames, <1 s) (Figure 3—figure supplement 2C3). They were not caused by confusion with other behaviors, such as wiping, grooming, rearing, locomotion, and resting (Figure 3—figure supplement 4). Type 1 error was also rare for manual annotation (Figure 3—figure supplement 3B). The models might miss short scratching trains, hence caused the Type 2 error. Indeed, all missed scratching trains were shorter than 40 frames (<1.3 s) (Figure 3—figure supplement 2D1, D2). For all scratching trains lasting less than 30 frames (<1 s), 18.5% of them were missed by the model prediction. This number decreased to 2.7% for scratching trains spanning between 30 and 60 frames (1–2 s). No scratching train was missed if they were longer than 60 frames (>2 s) (Figure 3—figure supplement 2D3). The Type 2 error positively correlated with the input length (N) of prediction models. It became zero or close to zero when models trained with shorter input lengths (3, 13, and 23 frames) (Figure 3—figure supplement 5A–C). For manual annotation, Type 2 error was not common (Figure 3—figure supplement 3C1, C2). Type 3 error (Figure 3—figure supplement 2E1–E3) was dominant among all five type errors. The average start and end frame shift of the model prediction were 2.2 and 7.0 frames (Figure 3—figure supplement 2E3), while those of the manual annotation were 11.5 and 12.8 frames (Figure 3—figure supplement 3D1, D2). The start and end frame shift of the manual quantification was similar (~350 ms), which likely reflected the temporal delay of the real time human visual system processing. For the model prediction, which was an off-line frame by frame process, the temporal shifts were less than the human visual processing. In addition, the start of a scratching train was more accurately recognized by the model than the end of a scratching train (Figure 3—figure supplement 2E3). This might reflect the feature of scratching trains. It was relatively clear to determine the start of a scratching train when a mouse lifted its hind paw, but more ambiguous to determine when a mouse put its hind paw back onto the floor to complete a scratching train. The start and end shift did not correlate with the length of a scratching train (Figure 3—figure supplement 5D). Thus, the relative error (percentage of error frames) would decrease when the duration of a scratching train increases. Indeed, the prediction accuracy (as indicated by the F1 score) positively correlated (R2 = 0.5723) with the average scratching train duration in a video (Figure 3—figure supplement 5E). Type 4 error was caused when two adjacent scratching trains were too close to each other and were predicted as one scratching train (Figure 3—figure supplement 2F1–F3). All missed intervals were shorter than 30 frames (<1 s) (Figure 3—figure supplement 2F2). Conversely, 51.4% of intervals less than 30 frames between the two adjacent scratching trains were not recognized. All intervals longer than 30 frames were recognized (Figure 3—figure supplement 2F3). Type 4 error was more common in manual annotation than in the model prediction (Figure 3—figure supplement 3E1, E2). Type 5 error occurred when one scratching train was predicted as two or more scratching trains, separated by mispredicted intervals. The average lengths of these mispredicted intervals were around 10 frames by model prediction and around 40 frames by manual annotation (Figure 3—figure supplement 2G1, G2 and Figure 3—figure supplement 3F). When reviewing these intervals, we found that more than 80% of them were within or partially overlapped with a paw licking phase (Figure 3—figure supplement 2G3), especially when the duration of the paw licking was more than 30 frames (Figure 3—figure supplement 5F). Thus, it seems likely that the model predicted some long licking frames within a scratching train as ‘non-scratching’. Type 4 and 5 errors reflect the intrinsic comple" @default.
W4311895184 created "2023-01-02" @default.
W4311895184 date "2022-11-29" @default.
W4311895184 modified "2023-09-27" @default.
W4311895184 title "Decision letter: Scratch-AID, a deep learning-based system for automatic detection of mouse scratching behavior with high accuracy" @default.
W4311895184 doi "https://doi.org/10.7554/elife.84042.sa1" @default.
W4311895184 hasPublicationYear "2022" @default.
W4311895184 type Work @default.
W4311895184 citedByCount "0" @default.
W4311895184 crossrefType "peer-review" @default.
W4311895184 hasBestOaLocation W43118951841 @default.
W4311895184 hasConcept C108583219 @default.
W4311895184 hasConcept C111919701 @default.
W4311895184 hasConcept C154945302 @default.
W4311895184 hasConcept C159985019 @default.
W4311895184 hasConcept C192562407 @default.
W4311895184 hasConcept C2780196728 @default.
W4311895184 hasConcept C2781235140 @default.
W4311895184 hasConcept C31972630 @default.
W4311895184 hasConcept C41008148 @default.
W4311895184 hasConceptScore W4311895184C108583219 @default.
W4311895184 hasConceptScore W4311895184C111919701 @default.
W4311895184 hasConceptScore W4311895184C154945302 @default.
W4311895184 hasConceptScore W4311895184C159985019 @default.
W4311895184 hasConceptScore W4311895184C192562407 @default.
W4311895184 hasConceptScore W4311895184C2780196728 @default.
W4311895184 hasConceptScore W4311895184C2781235140 @default.
W4311895184 hasConceptScore W4311895184C31972630 @default.
W4311895184 hasConceptScore W4311895184C41008148 @default.
W4311895184 hasLocation W43118951841 @default.
W4311895184 hasOpenAccess W4311895184 @default.
W4311895184 hasPrimaryLocation W43118951841 @default.
W4311895184 hasRelatedWork W1566683390 @default.
W4311895184 hasRelatedWork W1998037616 @default.
W4311895184 hasRelatedWork W2003287677 @default.
W4311895184 hasRelatedWork W2373778085 @default.
W4311895184 hasRelatedWork W2518838668 @default.
W4311895184 hasRelatedWork W2980116056 @default.
W4311895184 hasRelatedWork W4311944550 @default.
W4311895184 hasRelatedWork W4311944582 @default.
W4311895184 hasRelatedWork W4361273501 @default.
W4311895184 hasRelatedWork W602661801 @default.
W4311895184 isParatext "false" @default.
W4311895184 isRetracted "false" @default.
W4311895184 workType "peer-review" @default.