Highest F1 Score Research Articles

Software-dependent biases in the recognition of di- and tri-syllabic bird songs can create false interpretations of bird abundance and singing activity

The use of autonomous recording units (ARUs) for passive acoustic monitoring has recently gained a lot of importance. It is now widely used for scientific purposes and has been increasingly applied in the field of conservation biology. However, while ARUs may greatly increase cost-effectiveness in monitoring, they are not exempt from possible biases. In this study, we compared the performances of three different software in recognizing bi- and tri-syllabic bird songs and tested if their results varied significantly in a real setting. We focused on the interpretation of the relative abundance among habitats and the time of vocal activity, also using different numbers of training records to assess if the results of the software improved. Even with the quality measures of the recognizer being consistent with those of previously published studies, in several cases the results produced significantly different interpretations depending on the software used. While increasing the number of training records slightly improved the overlap in the estimates of activity among the different software, differences in the relative abundance of each song were present even with a fairly solid number of training records. Our results suggest that reliable biological interpretations might only be attained with a large number of training records and very high values of precision, recall, and the F-score. In fact, even with very high precision and F-scores, lower values for recall led not only to statistically significant differences, but also to different interpretations of relative abundance and time of activity between the two best-performing software. Thus, great caution must be taken in interpreting the results of bioacoustics studies that use automated recognition, and standards of minimum quality should be created for recognizers in scientific and technical studies.

Mechanical Equipment Failure Prediction Based on Machine Learning

Mechanical equipment plays a central role in the manufacturing industry. However, they often suffer from problems such as aging and wear, which pose a major challenge to maintaining the stability of the production environment and improving economic efficiency. In order to solve this problem, this article conducted an in-depth study using the usage and failure data of a certain enterprise's mechanical equipment. First, this article establishes an XGBoost classification model to determine whether mechanical equipment has malfunctioned. The model performed well, with performance metrics showing high precision, recall, and F1 scores. Furthermore, this paper used the XGBoost model to successfully predict the faults of 19 mechanical equipment, classified these devices, and accurately determined the types of faults, including heat dissipation faults, power faults and overload faults. Through the weight analysis of various types of fault data using the CRITIC weight analysis method, the main causes of various types of faults are clarified, which provides an important reference for further improvement and maintenance of this mechanical equipment.

Towards safer streets: A framework for unveiling pedestrians’ perceived road safety using street view imagery

Road safety has become a global concern but its impact in low- and middle-income countries is widespread mainly due to lack of appropriate crash database system and under-reporting. In this context, the primary objective of this paper is to provide a scalable framework for unveiling pedestrians’ perceived road safety that can also be applied in regions where accessible crash data are limited or near-crashes are left unreported. In the first step of our methodology, a deep learning architecture-based semantic segmentation model (HRNet+OCR) is trained using labeled Google Street View (GSV) images from specific study areas in Dhaka, Bangladesh, which facilitates the identification of both man-made components (such as roads, sidewalks, buildings, and vehicles) and natural elements (including trees and sky). The developed model showed excellent performance in identifying different features in an image by achieving high precision (0.95), recall (0.97), F1-score (0.96), and intersection over union (IoU) (91.86). Secondly, a group of trained raters scored the perceived road safety on an ordinal scale from 0 to 10 (extremely unsafe to extremely safe to walk in terms of road crashes) by assessing the GSV images. Then, several regression models have been used on features extracted from GSV images, and socio-demographic factors (i.e., population density, and relative wealth index) to estimate the perceived road safety, and random forest regression model was found to perform the best. Further, Shapley Additive Explanations (SHAP), a model-agnostic technique has been used for examining feature importance by computing the contribution of each feature to the random forest regression model output. The results show that sidewalk, road, population density, wall, and relative wealth index have higher impact on determining the perceived road safety rating. Additionally, the results of t-tests between the average perceived road safety scores for crash-prone and non crash-prone areas revealed the existence of significant differences. This study also provides perceived road safety rating map on a neighborhood scale, which can be a useful visualization tool for policy-makers and practitioners to identify the road safety deficiencies at specific locations, and formulate appropriate and strategic countermeasures to improve pedestrians’ road safety.

DCENSnet: A new deep convolutional ensemble network for skin cancer classification

Skin cancer is a significant health concern, demanding early detection and classification for higher survival rates. Existing computer vision methods struggle to tackle fine-grained variability in skin lesion features across different surfaces. Deep Convolutional Neural Network (DCNN) models show promise, but their current ad-hoc developments overlook redundant layers and suffer from imbalanced datasets and inadequate augmentation. To address these problems, we propose a novel ensemble approach involving three DCNNs, each customized with different dropout layers to enhance feature-level learning. Thus, the proposed ensemble network that we call DCENSnet achieves a superior bias–variance trade-off. Evaluating on the popular HAM10000 skin lesion dataset, our model outperforms state-of-the-art networks, achieving a mean accuracy of 99.53% along with high precision, recall, F1 score, and Area Under the ROC Curve (AUC) for each class. This method proves highly reliable for computer-aided detection, classification, and analysis of malignant skin lesions, holding promise for improving diagnosis and treatment accuracy.

Predicting Functional Outcome After Ischemic Stroke Using Logistic Regression and Machine Learning Models

This research employed binary logistic regression and machine learning techniques; Decision Tree, Random Forest, and Support Vector Machine (SVM), to predict functional outcomes following ischemic stroke. The main goal was to determine the most suitable model for the dataset through a comprehensive performance evaluation. Four models were examined for predicting post-ischemic stroke functional outcomes: Decision Tree, Random Forest, Logistic Regression, and SVM. The evaluation involved metrics such as Accuracy, Precision, F1-Score, and Recall. The Logistic Regression model achieved the highest accuracy at 90%, accurately predicting outcomes in 90% of cases. However, it had lower precision (50%), indicating an increased rate of false positive predictions. On the other hand, the SVM model displayed the highest precision (71.3%), implying fewer false positive predictions. It also attained the highest F1-Score (77.5%), indicating a strong balance between precision and Recall compared to the other models. Notably, the Logistic Regression model achieved perfect Recall (100%), correctly identifying all positive outcomes, while the Random Forest model showed significant recall performance (93.2%). Conversely, the Decision Tree model exhibited moderate accuracy (66.11%) but lower precision (66%), F1-Score (6.15%), and recall (3.2%), suggesting challenges with false positives and false negatives. Choosing the best model depends on analysis priorities. For accurate identification of positive outcomes, the Logistic Regression model's perfect recall is advantageous. For balanced performance, the SVM model's high F1-Score makes it a compelling option.

A lightweight deep learning based bowel sounds segmentation algorithm for gastrointestinal (GI) monitoring

Segmentation of bowel sounds (BS) events is a significant task of automatic BS monitoring. Recently, deep learning (DL) has been utilized to realize the segmentation of BS events. However, most researchers treat BS segmentation as a traditional classification problem, which causes that more precise locations for the occurrence of BS events are unable to obtained. Besides, the performance of segmentation of BS events is easily affected by thresholds of model output in practical applications. To tackle these issues, in this paper, a lightweight DL-based BS segmentation algorithm is proposed. The one-dimensional convolution layers and bidirectional gate recurrent unit (GRU) layers are adopted to enhance the ability of feature extraction. A loss function named shape loss is proposed to reduce the sensitivity of model to thresholds. Moreover, a portable BS monitoring device is developed to realize data acquisition, display and algorithm deployment. Based on this device, experiments on human and rats are conducted to verify the effectiveness of our proposed approach. Experimental results shows that the proposed method outperforms other comparison methods with the higher f1-scores and lower sensitivity to thresholds.

Detecting Fraudulent Transactions Using Stacked Autoencoder Kernel ELM Optimized by the Dandelion Algorithm

The risk of fraudulent activity has significantly increased with the rise in digital payments. To resolve this issue there is a need for reliable real-time fraud detection technologies. This research introduced an innovative method called stacked autoencoder kernel extreme learning machine optimized by the dandelion algorithm (S-AEKELM-DA) to detect fraudulent transactions. The primary objective was to enhance the kernel extreme learning machine (KELM) performance by integrating the dandelion technique into a stacked autoencoder kernel ELM architecture. This study aimed to improve the overall effectiveness of the proposed method in fraud detection by optimizing the regularization parameter (c) and the kernel parameter (σ). To evaluate the S-AEKELM-DA approach; simulations and experiments were conducted using four credit card datasets. The results demonstrated remarkable performance, with our method achieving high accuracy, recall, precision, and F1-score in real time for detecting fraudulent transactions. These findings highlight the effectiveness and reliability of the suggested approach. By incorporating the dandelion algorithm into the S-AEKELM framework, this research advances fraud detection capabilities, thus ensuring the security of digital transactions.

Learning to Segment Blob-like Objects by Image-Level Counting

There is a high demand for manually annotated data in many of the segmentation tasks based on neural networks. Selecting objects pixel by pixel not only takes much time, but it can also lead to inattentiveness and to inconsistencies due to changing annotators for different datasets and monotonous work. This is especially, but not exclusively, the case with sensor data such as microscopy imaging, where many blob-like objects need to be annotated. In addressing these problems, we present a weakly supervised training method that uses object counts at the image level to learn a segmentation implicitly instead of relying on a pixelwise annotation. Our method uses a given segmentation network and extends it with a counting head to enable training by counting. As part of the method, we introduce two specialized losses, contrast loss and morphological loss, which allow for a blob-like output with high contrast to be extracted from the last convolutional layer of the network before the actual counting. We show that similar high F1-scores can be achieved with weakly supervised learning methods as with strongly supervised training; in addition, we address the limitations of the presented method.

A fast, novel technique for detection of left ventricular hypertrophy by electrocardiography

Abstract Background Electrocardiography (ECG) is a widely used non-invasive tool for detecting left ventricular hypertrophy (LVH), but existing techniques suffer from complex rules and limited sensitivity. Purpose The goal of this study was to design and evaluate a new technique for rapid assessment of LVH by ECG. Methods The WS (Witteles-Somani) technique for LVH detection is based on recognizing the visual overlap between the S-wave of a precordial lead and the R-wave of any subsequent precordial lead. To evaluate its performance, ECGs from all patients at Stanford Healthcare with associated transthoracic echocardiograms (TTEs) between 2006 and 2018 were collected for analysis. ECGs with pre-existing bundle branch blocks or paced rhythms were excluded. A multiple adaptive thresholding technique was used to segment all QRS complexes from each ECG, with which algorithms to implement LVH detection criteria (i.e. WS, Sokolow-Lyon, Cornell, Peguero-Lo Presti) assessed performance over all ECGs. Associations between raw WS score (e.g. maximum S-wave and R-wave pair overlap value) and cardiovascular mortality, incident cardiovascular disease, and incident heart failure using age- and gender-adjusted Cox regression was also investigated. Results A total of 123,021 ECGs and 35,293 TTEs from 30,645 patients were collected. Of these ECGs, 36.1% belonged to patients with TTE-diagnosed LVH. The WS technique was more sensitive (46.1%) and had a higher F1-score (0.443) than other clinical criteria (Table 1), although with comparable calibration (AUROC 0.56 for WS vs. 0.54 for Sokolow-Lyon, 0.60 for Cornell, and 0.61 for Peguero-Lo Presti). Higher WS scores were associated with all-cause mortality (HR 1.08 [95% CI, 1.06, 1.11], P<0.005), incident cardiovascular disease (HR 1.15 [95% CI, 1.14, 1.17], P<0.005), and incident heart failure (HR 1.17 [95% CI, 1.16, 1.18], P<0.005). Conclusion The proposed WS technique is a much simpler and more sensitive technique than standard metrics to detect LVH from ECG, with important clinical implications in the development of future cardiovascular disease.Cumulative Incidence of DiseaseCriteria Classification Performance

Developing a Quality Control System in a Continuous Hot Air Heating Process in Surimi Seafood Processing Using Image Analysis and Artificial Intelligence

In this study, the feasibility of classifying surimi gels during a continuous heating process using an artificial intelligence (AI) algorithm on labeled images was investigated. Surimi paste with varying corn starch concentrations (0%, 5%, and 10%) and moisture content levels (78% and 80%) from Alaska pollock were analyzed for the subtle physical changes. Rheological characterization and K-means clustering analysis performed on entire images captured from different batches of heated surimi gel indicated a four-stage transformation from its initial state to gel formation with the temperature ranges spanning 25 to <40 °C, 40 to <50 °C, 50 to <55 °C, and 55 to 80 °C. Subsequently, a Convolutional Neural Network (CNN) model based on the temperature classification was designed to interpret and classify these images. A total of 1000 to 1200 images were used for the training, testing, and validation purposes in the ratio 7:1:2. The CNN architecture incorporated essential elements including an input layer, convolutional layers, rectified linear unit (ReLU) activation functions, normalization layers, and max-pooling layers. The CNN model achieved validation accuracy >92.67% for individual mixture composition, 94.53% for classifying surimi samples based on moisture content, and gelation level, and 89.73% for complex classifications involving moisture content, starch concentration, and gelation stages. Additionally, it exhibited high average precision, recall, and F1 scores (>0.92), indicating precision and sensitivity in identifying relevant instances. The success of CNN in non-destructively classifying surimi gels with different moisture and starch contents is demonstrated in this work.

A computer vision and residual neural network (ResNet) combined method for automated and accurate yeast replicative aging analysis of high-throughput microfluidic single-cell images

With the rapid development of microfluidic platforms in high-throughput single-cell culturing, laborious operation to manipulate massive budding yeast cells (Saccharomyces cerevisiae) in replicative aging studies has been greatly simplified and automated. As a result, large datasets of microscopy images bring challenges to fast and accurately determine yeast replicative lifespan (RLS), which is the most important parameter to study cell aging. Based on our microfluidic diploid yeast long-term culturing (DYLC) chip that features 1100 traps to immobilize single cells and record their proliferation and aging via time-lapse imaging, herein, a dedicated algorithm combined with computer vision and residual neural network (ResNet) was presented to efficiently process tremendous micrographs in a high-throughput and automated manner. The image-processing algorithm includes following pivotal steps: (i) segmenting multi-trap micrographs into time-lapse single-trap sub-images, (ii) labeling 8 yeast budding features and training the 18-layer ResNet, (iii) converting the ResNet predictions in analog values into digital signals, (iv) recognizing cell dynamic events, and (v) determining yeast RLS and budding time interval (BTI) ultimately. The ResNet algorithm achieved high F1 scores (over 92%) demonstrating the effectiveness and accuracy in the recognition of yeast budding events, such as bud appearance, daughter dissection and cell death. Therefore, the results conduct that similar deep learning algorithms could be tailored to analyze high-throughput microscopy images and extract multiple cell behaviors in microfluidic single-cell analysis.

Image patch-based deep learning approach for crop and weed recognition

Accurate classification of weed species in crop plants plays a crucial role in precision agriculture by enabling targeted treatment. Recent studies show that artificial intelligence deep learning (DL) models achieve promising solutions. However, several challenging issues, such as lack of adequate training data, inter-class similarity between weed species and intra-class dissimilarity between the images of the same weed species at different growth stages or for other reasons (e.g., variations in lighting conditions, image capturing mechanism, agricultural field environments) limit their performance. In this research, we propose an image based weed classification pipeline where a patch of the image is considered at a time to improve the performance. We first enhance the images using generative adversarial networks. The enhanced images are divided into overlapping patches, a subset of which are used for training the DL models. For selecting the most informative patches, we use the variance of Laplacian and the mean frequency of Fast Fourier Transforms. At test time, the model's outputs are fused using a weighted majority voting technique to infer the class label of an image. The proposed pipeline was evaluated using 10 state-of-the-art DL models on four publicly available crop weed datasets: DeepWeeds, Cotton weed, Corn weed, and Cotton Tomato weed. Our pipeline achieved significant performance improvements on all four datasets. DenseNet201 achieved the top performance with F1 scores of 98.49%, 99.83% and 100% on Deepweeds, Corn weed and Cotton Tomato weed datasets, respectively. The highest F1 score on the Cotton weed dataset was 98.96%, obtained by InceptionResNetV2. Moreover, the proposed pipeline addressed the issues of intra-class dissimilarity and inter-class similarity in the DeepWeeds dataset and more accurately classified the minority weed classes in the Cotton weed dataset. This performance indicates that the proposed pipeline can be used in farming applications.

Leveraging a Hybrid Deep Learning Architecture for Efficient Emotion Recognition in Audio Processing

This paper presents a novel hybrid deep learning architecture for emotion recognition from speech signals, which has garnered significant interest in recent years due to its potential applications in various fields such as healthcare, psychology, and entertainment. The proposed architecture combines modified ResNet-34 and RoBERTa models to extract meaningful features from speech signals and classify them into different emotion categories. The model is evaluated on five standard emotion recognition datasets, including RAVDESS, EmoDB, SAVEE, CREMA-D, and TESS, and achieves state-of-the-art performance on all datasets. The experimental results show that the proposed hybrid architecture outperforms existing emotion recognition models, achieving high accuracy and F1 scores for emotion classification. The proposed architecture is promising for real-time emotion recognition applications and can be applied in various domains such as speech-based emotion recognition systems, human-computer interaction, and virtual assistants.

Clinical Text Reports to Stratify Patients Affected with Myeloid Neoplasms Using Natural Language Processing

Background: The availability of multimodal patient data, such as demographics, clinical, imaging, treatment, quality of life, outcomes and wearables data, as well as genome sequencing, have paved the way for the development of multimodal clinical solutions that introduce personalized or precision medicine. The clinical report is an information layer that contains relevant information about the disease in addition to the patient's point of view. Natural language processing (NLP) is a branch of artificial intelligence (AI) and its pre-trained language models are the key technology for extracting value from this data layer. Aims: This project was conducted by GenoMed4all and Synthema EU consortia, with the aim to: 1) Build an AI language model specific for the hematology domain. 2) Use NLP technology to extract relevant information from clinical reports and perform unsupervised stratification of patients, in order to 3) demonstrate that the clinical report is earlier access to data relative to disease clinical phenotype and biology and provide important information for patient stratification and prediction of clinical outcomes. Methods: To translate text sentences into numerical embeddings, we implemented bidirectional encoder representations from transformers (BERT) framework. To learn text representations and correlations within data, we performed domain-adaptation by fine-tuned pre-trained model on hematological clinical reports of patients with myeloproliferative neoplasms (MPN), myelodysplastic syndrome (MDS) and acute myeloid leukemia (AML). Patient stratification was performed by HDBSCAN clustering on text embedding encoded by BERT (HematoBERT). Clusters validation was performed by assessing patients' diagnosis and survival probability. Finally, we compared domain-tuned HematoBERT vs pre-trained non-contextualized models. Results: We implemented HematoBERT based on the bert-base-multilingual-uncased version of BERT. Training data were hematological text reports of 1,328 patients. During fine-tuning, texts were tokenized, then we randomly replaced 15% of the tokens with masked tokens, training the model to predict them. We performed stratification using clinical reports from a validation cohort of 360 patients. We identified 7 clusters, defined according to similar words in meaning that were placed in a specific topic. We extracted the most important words and concepts for each cluster (topic) and we summarized them into effective descriptions for each group of patients. Two clusters included MDS patients with excess blasts, and without excess blasts with ring sideroblasts and del5q (n=69, n=115). One cluster included patients with excess blasts and MDS/MPN (n= 33). Two clusters included MPN patients with primary and secondary myelofibrosis, and MPN patients most including subjects affected with polycythemia vera and essential thrombocythemia (n=35, n=46). Two clusters included patients with AML from MDS and therapy-related AML, and patients with de novo AML (n=22, n=42). Clinical validation was performed based on the diagnosis and survival probability of patients assigned to clusters. Patients' diagnoses were compatible with the cluster assignment (Figure 1). Frequency of gene mutations (as assessed by targeted Next-Generation Sequencing) among different clusters reflected the well-known genotypic-phenotypic associations in MDS, MPN and AML. Kaplan-Maier curves indicated significative risk stratification in clusters in terms of survival probability (Figure 2), similar to stratifications performed on clinical and genomic data. Finally, we evaluate the domain adaptation by comparing the model to other pre-trained non-contextualized ones. Pseudo perplexity score (PPS), accuracy and F1 score were calculated to quantify how good the models are when they see new data, predicting the next word given the context of the sentence. HematoBERT obtained high PPS, accuracy and F1 scores, outperforming the other models also trained on generic clinical domains. Conclusion: Domain-adapted language models are able to understand contexts and correlations in documents. HematoBERT can be used to extract relevant features from clinical reports. This data layer is relevant to perform disease stratification of patients based on clinical and genomic information and could be integrated into next-generation multimodal models of personalized medicine.

A Powerful Correspondence Selection Method for Point Cloud Registration Based on Machine Learning

Correspondence selection is an indispensable process in point cloud registration. The success of point cloud registration largely depends on a good correspondence selection method. For this purpose, a novel correspondence selection method is proposed in this paper. First, two geometric constraints, one of which is proposed in this paper, are used to compute the compatibility score between two correspondences. Then, the feature vectors of the correspondences are constructed according to the compatibility scores between the correspondence and others. A support vector machine classifier is trained to classify the correct and incorrect correspondences by using the feature vectors. The experimental results demonstrate that our method can choose the right correspondences well and get high precision and F-score performance. Also, our method has the best robustness to noise, pointdensity variation, and partial overlap compared to the other methods.

Enhancing prediction of tooth caries using significant features and multi-model classifier

Background Tooth decay, also known as dental caries, is a common oral health problem that requires early diagnosis and treatment to prevent further complications. It is a chronic disease that causes the gradual breakdown of the tooth’s hard tissues, primarily due to the interaction of bacteria and dietary sugars. Results While numerous investigations have focused on addressing this issue using image-based datasets, the outcomes have revealed limitations in their effectiveness. In a novel approach, this study focuses on feature-based datasets, coupled with the strategic integration of Principle Component Analysis (PCA) and Chi-square (chi2) for robust feature engineering. In the proposed model, features are generated using PCA, utilizing a voting classifier ensemble consisting of Extreme Gradient Boosting (XGB), Random Forest (RF), and Extra Trees Classifier (ETC) algorithms. Discussion Extensive experiments were conducted to compare the proposed approach with the chi2 features and machine learning models to evaluate its efficacy for tooth caries detection. The results showed that the proposed voting classifier using PCA features outperformed the other approaches, achieving an accuracy, precision, recall, and F1 score of 97.36%, 96.14%, 96.84%, and 96.65%, respectively. Conclusion The study demonstrates that the utilization of feature-based datasets and PCA-based feature engineering, along with a voting classifier ensemble, significantly improves tooth caries detection accuracy compared to image-based approaches. The achieved high accuracy, precision, recall, and F1 score emphasize the potential of the proposed model for effective dental caries detection. This study provides new insights into the potential of innovative methodologies to improve dental healthcare by evaluating their effectiveness in addressing prevalent oral health issues.

A Survey on HealthCare Claim Prediction

Abstract: The life insurance industry has experienced a significant increase in claims in recent years. To cope with the large volume of claims, insurance companies are turning to machine learning models for predicting life insurance claims. This paper presents a machine-learning model for predicting life insurance claims based on various demographic and health-related features. The proposed model will be evaluated on a real dataset and will achieve a high accuracy, precision, recall, and F1 score. The results will show that the machine learning model can effectively predict life insurance claims, which can improve the efficiency of claims processing and ultimately benefit insurance companies and policyholders

Online multi-objective optimization for tunnel boring machine segment assembly considering stress concentration

To address stress concentration issues during the assembly of tunnel boring machine (TBM) segments, particularly in the context of synchronous excavation and segment assembly (SESA) TBMs, this paper presents an optimization framework focused on hydraulic cylinder pressure redistribution. The paper develops a Bayesian optimization extreme gradient boosting (BO-XGBoost) model for cylinder status detection and devises an online Archimedean optimization algorithm (AOA) for cylinder pressure redistribution. Verification of the proposed approach's effectiveness is conducted using a SESA TBM experimental bench. The results indicate that: 1) The proposed hydraulic cylinder status detection model effectively distinguishes various cylinder operating statuses, achieving a high F1-score of 0.9891. 2) Across the six segment assembly scenarios, the stress concentration phenomenon can be improved by 16.03%, 14.29%, 8.84%, 31.10%, 15.47%, and 11.11%, respectively. 3) In comparison to the offline optimization algorithm, the online algorithm reduces the maximum pressure by 36.69% and decreases algorithmic error by 30.50%. A key contribution from this paper is the introduction of an active pressure control approach integrated with the online multi-objective optimization algorithm, which is helpful in effectively improving the stress concentration phenomenon in the SESA TBM.

Artificial intelligence classification and amylose content prediction of rice flour varieties from their pasting features

The pasting features of eight varieties of rice flour with varying amylose content were characterized and then subjected to machine learning analysis in order to classify the rice varieties and predict their amylose content. The waxy rice flour with the least amylose showed a peak viscosity in the early stage of heating, and the high amylose rice flours had the highest final viscosity. Principal component analysis showed that 87.2% of the total variability was explained by the two principal components, which were mainly related to peak temperature and final viscosity. When the pasting features were used as a machine learning dataset for classification, the support vector machine classifier was the most effective in correctly classifying the rice flour varieties by showing a high accuracy and F1-score, followed by the decision tree and stochastic gradient descent. In addition, the integration of pasting features with machine learning analysis showed the potential to predict the amylose content of rice flours. These prediction performances were confirmed by validating the models with independent datasets.

Adaptive undersampling and short clip-based two-stream CNN-LSTM model for surgical phase recognition on cholecystectomy videos

Surgical phase recognition is challenging due to overfitting problems caused by imbalanced data among surgical phases. We proposed an adaptive sampling rate-based undersampling method that could generate the number of each surgical phase data similarly to alleviate biased learning. To improve the performance of our method, we also introduced a two-stream CNN-LSTM model that could extract temporal information on behavioral changes between each image frame. First, we extracted a total of 40,236 short clips using an adaptive subsampling rate from the entire video. Each short clip was entered into a pre-trained GoogLeNet. The output with visual information was then immediately fed into a sequence-to-sequence LSTM model to extract temporal information of neighbor frames within a short clip. At the same time, another sequence-to-vector LSTM was used, to extract temporal information from all successive image frames to predict the final surgical phase. The proposed method was evaluated with a public dataset Cholec80. The proposed approach outperformed state-of-the-art methods, showing a high F1-score of 87.12% and an AUC of 98.00%. In addition, the F1-score deviation between all phases decreased by about 10% compared to that before applying undersampling. Experimental results confirmed that employing our proposed method could learn enrich temporal information from short clips. It outperformed the conventional one-stream CNN-LSTM architecture.