Deep-learning seismology

DeepUSPS: Deep Learning-Empowered Unconstrained-Structural Protein Sequence Design

Currently, the unconstrained-structural protein sequence design models suffer from low optimization efficiency, and their generated proteins exhibit significant similarities to natural proteins and low thermal stability. To address these challenges, we propose the Deep Learning-Empowered Unconstrained-Structural Protein Sequence Design (DeepUSPS) model. To effectively address the inadequate thermal stability problem, we employ the innovative Inverted Dense Residual Network (IDRNet). To mitigate the designed proteins similarity issue, the Sequence-Pairwise Features Extraction Synthetic Network (SPFESN) is constructed. Furthermore, we introduce the Warm Restart AngularGrad (WRA) optimizer to optimize the 3D Position-Specific Scoring Matrix (3Dpssm) for unconstrained-structural protein sequence, only involving 2100 iterations (140.36 min) updates to generate idealization (IDE) protein sequences. We obtained a total of 1000 IDE protein sequences. Then we utilized in silico experiments to evaluate them, including similarity, clarity and iterations, thermal stability, spatial distribution of similarity, and predicted local-distance difference test (pLDDT) confidence assessment. Notably, the mean lg(E-value) for IDE protein sequences reached -0.051, the mean TM-score for IDE protein structures reached 0.594, the iterations only need 2100, and the mean Tm (melting point) for thermal stability reached 74.78°C. The average pLDDT value for 3D structures reached 76. Additionally, the IDE proteins' 3D structures exhibit diverse types. These in silico results conclusively demonstrate the superior performance of DeepUSPS compared with Hallucinate.

Predicting bone metastasis risk of colorectal tumors using radiomics and deep learning ViT model

Background

Colorectal cancer is a prevalent malignancy with a significant risk of metastasis, including to bones, which severely impacts patient outcomes. Accurate prediction of bone metastasis risk is crucial for optimizing treatment strategies and improving prognosis.

Purpose

This study aims to develop a predictive model combining radiomics and Vision Transformer (ViT) deep learning techniques to assess the risk of bone metastasis in colorectal cancer patients using both plain and contrast-enhanced CT images.

Materials and methods

We conducted a retrospective analysis of 155 colorectal cancer patients, including 81 with bone metastasis and 74 without. Radiomic features were extracted from segmented tumors on both plain and contrast-enhanced CT images. LASSO regression was applied to select key features, which were then used to build traditional machine learning models, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest, LightGBM, and XGBoost. Additionally, a dual-modality ViT model was trained on the same CT images, with a late fusion strategy employed to combine outputs from the different modalities. Model performance was evaluated using AUC-ROC, accuracy, sensitivity, and specificity, and differences were statistically assessed using DeLong's test.

Results

The ViT model demonstrated superior predictive performance, achieving an AUC of 0.918 on the test set, significantly outperforming all traditional radiomics-based models. The SVM model, while the best among traditional models, still underperformed compared to the ViT model. The ViT model's strength lies in its ability to capture complex spatial relationships and long-range dependencies within the imaging data, which are often missed by traditional models. DeLong's test confirmed the statistical significance of the ViT model's enhanced performance, highlighting its potential as a powerful tool for predicting bone metastasis risk in colorectal cancer patients.

Conclusion

The integration of radiomics with ViT-based deep learning offers a robust and accurate method for predicting bone metastasis risk in colorectal cancer patients. The ViT model's ability to analyze dual-modality CT imaging data provides greater precision in risk assessment, which can improve clinical decision-making and personalized treatment strategies. These findings underscore the promise of advanced deep learning models in enhancing the accuracy of metastasis prediction. Further validation in larger, multicenter studies is recommended to confirm the generalizability of these results.

Comprehensive SHAP Values and Single-Cell Sequencing Technology Reveal Key Cell Clusters in Bovine Skeletal Muscle

The skeletal muscle of cattle is the main component of their muscular system, responsible for supporting and movement functions. However, there are still many unknown areas regarding the ranking of the importance of different types of cell populations within it. This study conducted in-depth research and made a series of significant findings. First, we trained 15 bovine skeletal muscle models and selected the best-performing model as the initial model. Based on the SHAP (Shapley Additive exPlanations) analysis of this initial model, we obtained the SHAP values of 476 important genes. Using the contributions of these 476 genes, we reconstructed a 476-gene SHAP value matrix, and relying solely on the interactions among these 476 genes, successfully mapped the single-cell atlas of bovine skeletal muscle. After retraining the model and further interpretation, we found that Myofiber cells are the most representative cell type in bovine skeletal muscle, followed by neutrophils. By determining the key genes of each cell type through SHAP values, we conducted analyses on the correlations among key genes and between cells for Myofiber cells, revealing the critical role these genes play in muscle growth and development. Further, by using protein language models, we performed cross-species comparisons between cattle and pigs, deepening our understanding of Myofiber cells as key cells in skeletal muscle, and exploring the common regulatory mechanisms of muscle development across species.

Low-frequency tremor-like episodes before the 2023 MW 7.8 Türkiye earthquake linked to cement quarrying

Recent advances in artificial intelligence have enhanced the detection and identification of transient low-amplitude signals across the entire frequency spectrum, shedding light on deformation processes preceding natural hazards. This study investigates low-frequency, low-amplitude signals preceding the 2023 MW 7.8 Kahramanmaraş earthquake in Türkiye. Using a deep neural network, we extract key features from the spectrograms of continuous seismic signals and employ unsupervised clustering to reveal distinct transient patterns. We identify an increased occurrence of low-frequency tremor-like signals during the six months preceding the mainshock. However, the location of these signals suggests that their origin is not tectonic, but rather related to anthropogenic activities at cement plants along the Narlı Fault, where the MW 7.8 mainshock nucleated. Such findings highlight the importance of understanding the origin of patterns detected by machine-learning methods and the large variety of seismic signals due to anthropogenic activities. Furthermore, the search for the origin of the tremor-like signals motivated an investigation into the local seismicity around the Narlı Fault. The resulting extended seismicity catalog suggests that seismicity in this area arises from a combination of tectonic and anthropogenic processes.

Diagnosis of clear cell renal cell carcinoma via a deep learning model with whole-slide images

Background

Traditional pathological diagnosis methods have limitations in terms of interobserver variability and the time consumption of evaluations. In this study, we explored the feasibility of using whole-slide images (WSIs) to establish a deep learning model for the diagnosis of clear cell renal cell carcinoma (ccRCC).

Methods

We retrospectively collected pathological data from 95 patients with ccRCC from January 2023 to December 2023. All pathological slices conforming to the standards of the model were manually annotated first. The WSIs were preprocessed to extract the region of interest. The WSIs were divided into a training set and a test set, and the ratio of tumor slices to normal tissue slices in the training set to the test set was 3:1. Positive and negative samples were randomly extracted. Model training was based on a convolutional neural network (CNN) and a random forest model. The accuracy of the model was evaluated by generating a receiver operating characteristic (ROC) curve.

Results

A total of 663 pathological slices from 95 patients with ccRCC were collected. The mean number of slices per patient was 7.6 ± 2.7 (range: 3-17), with 506 tumor slices and 157 normal tissue slices. There were 200 tumor slices and 74 normal slices in the training set, and a total of 200,870 small images were extracted. There were 250 tumor slices and 63 normal slices in the test set, and a total of 39,211 small images were extracted. According to the CNN model and random forest model trained with the training set, 11 pathological slices in the test set were identified as false normal slices, and six pathological slices were identified as false tumor slices. The total accuracy was 94.6% (296/313), the precision rate was 97.6% (239/245), and the recall rate was 95.6% (239/250). The generated probabilistic heatmaps were consistent with the manually annotated pathological images. The ROC curve results revealed that the area under curve (AUC) reached 0.9658 (95% confidence interval: 0.9603-0.9713), the specificity was 90.5%, and the sensitivity was 95.6%.

Conclusion

The use of a deep learning method for the diagnosis of ccRCC is feasible. The ccRCC model established in this study achieved high accuracy. AI-based diagnostic methods for ccRCC may improve diagnostic efficiency.

Pattern memory cannot be completely and truly realized in deep neural networks

The unknown boundary issue, between superior computational capability of deep neural networks (DNNs) and human cognitive ability, has becoming crucial and foundational theoretical problem in AI evolution. Undoubtedly, DNN-empowered AI capability is increasingly surpassing human intelligence in handling general intelligent tasks. However, the absence of DNN's interpretability and recurrent erratic behavior remain incontrovertible facts. Inspired by perceptual characteristics of human vision on optical illusions, we propose a novel working capability analysis framework for DNNs through innovative cognitive response characteristics on visual illusion images, accompanied with fine adjustable sample image construction strategy. Our findings indicate that, although DNNs can infinitely approximate human-provided empirical standards in pattern classification, object detection and semantic segmentation, they are still unable to truly realize independent pattern memorization. All super cognitive abilities of DNNs purely come from their powerful sample classification performance on similar known scenes. Above discovery establishes a new foundation for advancing artificial general intelligence.

Discrimination of doubled Acoustic Emission events using neural networks

In observatory seismology, the effective automatic processing of seismograms is a time-consuming task. A contemporary approach for seismogram processing is based on the Deep Neural Network formalism, which has been successfully applied in many fields. Here, we present a 4D network, based on U-net architecture, that simultaneously processes seismograms from an entire network. We also interpret Acoustic Emission data based on a laboratory loading experiment. The obtained data was a very good testing set, similar to real seismograms. Our Neural network is designed to detect multiple events. Input data are created by augmentation from previously interpreted single events. The advantage of the approach is that the positions of (multiple) events are exactly known, thus, the efficiency of detection can be evaluated. Even if the method reaches an average efficiency of only around 30% for the onset of individual tracks, average efficiency for the detection of double events was approximately 97% for a maximum target, with a prediction difference of 20 samples. Such is the main benefit of simultaneous network signal processing.

The advantages of artificial intelligence-assisted total hip arthroplasty: A randomized controlled trial followed by 12 months

OBJECTION

The rapid advancement of artificial intelligence has brought significant breakthroughs to various medical disciplines，This study aimed to compare perioperative factors and postoperative hip function recovery in primary total hip arthroplasty (THA)by evaluating the use of an artificial intelligence (AI) preoperative planning system versus traditional two-dimensional X-ray planning.

PATIENTS AND METHODS

A total of 45 eligible patients underwent primary THA at Beijing Shijitan Hospital, Capital Medical University, between July 2022 and August 2022. The patients were randomly assigned to either the experimental group (n = 19) or the control group (n = 26). The experimental group utilized AI planning, while the control group employed traditional two-dimensional X-ray planning. Statistical analysis was performed to compare the accuracy of prosthesis prediction, operation time, intraoperative blood loss, frequency of intraoperative model testing, length of hospital stay, postoperative imaging data, and postoperative hip function scores. These comparisons were made to assess the effects of different preoperative planning methods on perioperative and postoperative hip function recovery.

RESULTS

The accuracy of preoperative planning for the acetabular and femoral sides in AI-assisted total hip arthroplasty was 84.2 % and 89.5 %, respectively, which was significantly better than that of the traditional two-dimensional X-ray planning group (P < 0.05). The operation time for AI-assisted total hip arthroplasty was 104.32 ± 18.10 min, which was shorter than that of the traditional two-dimensional X-ray planning group (P < 0.05). At 3 months post-operation, the grade of Harris score for hip function in the AI planning group was significantly better than that in the traditional two-dimensional X-ray planning group (P < 0.05). The average postoperative Harris score of the artificial intelligence group was higher than that of the traditional two-dimensional X-ray planning (P < 0.05).

CONCLUSION

Artificial intelligence-assisted total hip arthroplasty demonstrated superior accuracy in prosthesis prediction, shorter operation time, higher average Harris score at postoperative follow-up, and better hip function recovery at 3 months compared to traditional two-dimensional X-ray planning.

LEVEL OF EVIDENCE

III, case-control study.

Deep bone oncology Diagnostics: Computed tomography based Machine learning for detection of bone tumors from breast cancer metastasis

Purpose

The objective of this study is to develop a novel diagnostic tool using deep learning and radiomics to distinguish bone tumors on CT images as metastases from breast cancer. By providing a more accurate and reliable method for identifying metastatic bone tumors, this approach aims to significantly improve clinical decision-making and patient management in the context of breast cancer.

Methods

This study utilized CT images of bone tumors from 178 patients, including 78 cases of breast cancer bone metastases and 100 cases of non-breast cancer bone metastases. The dataset was processed using the Medical Image Segmentation via Self-distilling TransUNet (MISSU) model for automated segmentation. Radiomics features were extracted from the segmented tumor regions using the Pyradiomics library, capturing various aspects of tumor phenotype. Feature selection was conducted using LASSO regression to identify the most predictive features. The model's performance was evaluated using ten-fold cross-validation, with metrics including accuracy, sensitivity, specificity, and the Dice similarity coefficient.

Results

The developed radiomics model using the SVM algorithm achieved high discriminatory power, with an AUC of 0.936 on the training set and 0.953 on the test set. The model's performance metrics demonstrated strong accuracy, sensitivity, and specificity. Specifically, the accuracy was 0.864 for the training set and 0.853 for the test set. Sensitivity values were 0.838 and 0.789 for the training and test sets, respectively, while specificity values were 0.896 and 0.933 for the training and test sets, respectively. These results indicate that the SVM model effectively distinguishes between bone metastases originating from breast cancer and other origins. Additionally, the average Dice similarity coefficient for the automated segmentation was 0.915, demonstrating a high level of agreement with manual segmentations.

Conclusion

This study demonstrates the potential of combining CT-based radiomics and deep learning for the accurate detection of bone metastases from breast cancer. The high-performance metrics indicate that this approach can significantly enhance diagnostic accuracy, aiding in early detection and improving patient outcomes. Future research should focus on validating these findings on larger datasets, integrating the model into clinical workflows, and exploring its use in personalized treatment planning.

Self-evolving artificial intelligence framework to better decipher short-term large earthquakes

Large earthquakes (EQs) occur at surprising loci and timing, and their descriptions remain a long-standing enigma. Finding answers by traditional approaches or recently emerging machine learning (ML)-driven approaches is formidably difficult due to data scarcity, interwoven multiple physics, and absent first principles. This paper develops a novel artificial intelligence (AI) framework that can transform raw observational EQ data into ML-friendly new features via basic physics and mathematics and that can self-evolve in a direction to better reproduce short-term large EQs. An advanced reinforcement learning (RL) architecture is placed at the highest level to achieve self-evolution. It incorporates transparent ML models to reproduce magnitude and spatial location of large EQs ([Formula: see text] 6.5) weeks before of the failure. Verifications with 40-year EQs in the western U.S. and comparisons against a popular EQ forecasting method are promising. This work will add a new dimension of AI technologies to large EQ research. The developed AI framework will help establish a new database of all EQs in terms of ML-friendly new features and continue to self-evolve in a direction of better reproducing large EQs.

Accuracy of artificial intelligence-assisted endoscopy in the diagnosis of gastric intestinal metaplasia: A systematic review and meta-analysis

BACKGROUND AND AIMS

Gastric intestinal metaplasia is a precancerous disease, and a timely diagnosis is essential to delay or halt cancer progression. Artificial intelligence (AI) has found widespread application in the field of disease diagnosis. This study aimed to conduct a comprehensive evaluation of AI's diagnostic accuracy in detecting gastric intestinal metaplasia in endoscopy, compare it to endoscopists' ability, and explore the main factors affecting AI's performance.

METHODS

The study followed the PRISMA-DTA guidelines, and the PubMed, Embase, Web of Science, Cochrane, and IEEE Xplore databases were searched to include relevant studies published by October 2023. We extracted the key features and experimental data of each study and combined the sensitivity and specificity metrics by meta-analysis. We then compared the diagnostic ability of the AI versus the endoscopists using the same test data.

RESULTS

Twelve studies with 11,173 patients were included, demonstrating AI models' efficacy in diagnosing gastric intestinal metaplasia. The meta-analysis yielded a pooled sensitivity of 94% (95% confidence interval: 0.92-0.96) and specificity of 93% (95% confidence interval: 0.89-0.95). The combined area under the receiver operating characteristics curve was 0.97. The results of meta-regression and subgroup analysis showed that factors such as study design, endoscopy type, number of training images, and algorithm had a significant effect on the diagnostic performance of AI. The AI exhibited a higher diagnostic capacity than endoscopists (sensitivity: 95% vs. 79%).

CONCLUSIONS

AI-aided diagnosis of gastric intestinal metaplasia using endoscopy showed high performance and clinical diagnostic value. However, further prospective studies are required to validate these findings.

A review of the application of deep learning in obesity: From early prediction aid to advanced management assistance

BACKGROUND AND AIMS

Obesity is a chronic disease which can cause severe metabolic disorders. Machine learning (ML) techniques, especially deep learning (DL), have proven to be useful in obesity research. However, there is a dearth of systematic reviews of DL applications in obesity. This article aims to summarize the current trend of DL usage in obesity research.

METHODS

An extensive literature review was carried out across multiple databases, including PubMed, Embase, Web of Science, Scopus, and Medline, to collate relevant studies published from January 2018 to September 2023. The focus was on research detailing the application of DL in the context of obesity. We have distilled critical insights pertaining to the utilized learning models, encompassing aspects of their development, principal results, and foundational methodologies.

RESULTS

Our analysis culminated in the synthesis of new knowledge regarding the application of DL in the context of obesity. Finally, 40 research articles were included. The final collection of these research can be divided into three categories: obesity prediction (n = 16); obesity management (n = 13); and body fat estimation (n = 11).

CONCLUSIONS

This is the first review to examine DL applications in obesity. It reveals DL's superiority in obesity prediction over traditional ML methods, showing promise for multi-omics research. DL also innovates in obesity management through diet, fitness, and environmental analyses. Additionally, DL improves body fat estimation, offering affordable and precise monitoring tools. The study is registered with PROSPERO (ID: CRD42023475159).

Identification of key immune cells infiltrated in lung adenocarcinoma microenvironment and their related long noncoding RNA

LncRNA associated with immune cell infiltration in tumor microenvironment (TME) may be a potential therapeutic target for lung adenocarcinoma. We established a machine learning (ML) model based on 3896 samples characterized by the degree of immune cell infiltration, and further screened the key lncRNA. In vitro experiments were applied to validate the prediction. Treg is the key immune cell in the TME of lung adenocarcinoma, and the degree of infiltration is negatively correlated with the prognosis. PCBP1-AS1 may affect the infiltration of Tregs by regulating the TGF-β pathway, which is a potential predictor of clinical response to immunotherapy. PCBP1-AS1 regulates cell proliferation, cell cycle, invasion, migration, and apoptosis in lung adenocarcinoma. The results of clinical sample staining and in vitro experiments showed that PCBP1-AS1 was negatively correlated with Treg infiltration and TGF-β expression. Tregs and related lncRNA PCBP1-AS1 can be used as targets for the diagnosis and treatment of lung adenocarcinoma.

Using Fuzzy C-Means Clustering to Determine First Arrival of Microseismic Recordings

Accurate and automatic first-arrival picking is one of the most crucial steps in microseismic monitoring. We propose a method based on fuzzy c-means clustering (FCC) to accurately divide microseismic data into useful waveform and noise sections. The microseismic recordings' polarization linearity, variance, and energy are employed as inputs for the fuzzy clustering algorithm. The FCC produces a membership degree matrix that calculates the membership degree of each feature belonging to each cluster. The data section with the higher membership degree is identified as the useful waveform section, whose first point is determined as the first arrival. The extracted polarization linearity improves the classification performance of the fuzzy clustering algorithm, thereby enhancing the accuracy of first-arrival picking. Comparison tests using synthetic data with different signal-to-noise ratios (SNRs) demonstrate that the proposed method ensures that 94.3% of the first arrivals picked have an error within 2 ms when SNR = -5 dB, surpassing the residual U-Net, Akaike information criterion, and short/long time average ratio approaches. In addition, the proposed method achieves a picking accuracy of over 95% in the real dataset tests without requiring labelled data.

Virtual Scenarios of Earthquake Early Warning to Disaster Management in Smart Cities Based on Auxiliary Classifier Generative Adversarial Networks

Effective response strategies to earthquake disasters are crucial for disaster management in smart cities. However, in regions where earthquakes do not occur frequently, model construction may be difficult due to a lack of training data. To address this issue, there is a need for technology that can generate earthquake scenarios for response training at any location. We proposed a model for generating earthquake scenarios using an auxiliary classifier Generative Adversarial Network (AC-GAN)-based data synthesis. The proposed ACGAN model generates various earthquake scenarios by incorporating an auxiliary classifier learning process into the discriminator of GAN. Our results at borehole sensors showed that the seismic data generated by the proposed model had similar characteristics to actual data. To further validate our results, we compared the generated IM (such as PGA, PGV, and SA) with Ground Motion Prediction Equations (GMPE). Furthermore, we evaluated the potential of using the generated scenarios for earthquake early warning training. The proposed model and algorithm have significant potential in advancing seismic analysis and detection management systems, and also contribute to disaster management.

Arrival times by Recurrent Neural Network for induced seismic events from a permanent network

We have developed a Recurrent Neural Network (RNN)-based phase picker for data obtained from a local seismic monitoring array specifically designated for induced seismicity analysis. The proposed algorithm was rigorously tested using real-world data from a network encompassing nine three-component stations. The algorithm is designed for multiple monitoring of repeated injection within the permanent array. For such an array, the RNN is initially trained on a foundational dataset, enabling the trained algorithm to accurately identify other induced events even if they occur in different regions of the array. Our RNN-based phase picker achieved an accuracy exceeding 80% for arrival time picking when compared to precise manual picking techniques. However, the event locations (based on the arrival picking) had to be further constrained to avoid false arrival picks. By utilizing these refined arrival times, we were able to locate seismic events and assess their magnitudes. The magnitudes of events processed automatically exhibited a discrepancy of up to 0.3 when juxtaposed with those derived from manual processing. Importantly, the efficacy of our results remains consistent irrespective of the specific training dataset employed, provided that the dataset originates from within the network.

Fast and Efficient Design of Deep Neural Networks for Predicting N7-Methylguanosine Sites Using autoBioSeqpy

N⁷-Methylguanosine (m⁷G) is a crucial post-transcriptional RNA modification that plays a pivotal role in regulating gene expression. Accurately identifying m⁷G sites is a fundamental step in understanding the biological functions and regulatory mechanisms associated with this modification. While whole-genome sequencing is the gold standard for RNA modification site detection, it is a time-consuming, expensive, and intricate process. Recently, computational approaches, especially deep learning (DL) techniques, have gained popularity in achieving this objective. Convolutional neural networks and recurrent neural networks are examples of DL algorithms that have emerged as versatile tools for modeling biological sequence data. However, developing an efficient network architecture with superior performance remains a challenging task, requiring significant expertise, time, and effort. To address this, we previously introduced a tool called autoBioSeqpy, which streamlines the design and implementation of DL networks for biological sequence classification. In this study, we utilized autoBioSeqpy to develop, train, evaluate, and fine-tune sequence-level DL models for predicting m⁷G sites. We provided detailed descriptions of these models, along with a step-by-step guide on their execution. The same methodology can be applied to other systems dealing with similar biological questions. The benchmark data and code utilized in this study can be accessed for free at http://github.com/jingry/autoBioSeeqpy/tree/2.0/examples/m7G.

Encoder-Decoder Architecture for 3D Seismic Inversion

Inverting seismic data to build 3D geological structures is a challenging task due to the overwhelming amount of acquired seismic data, and the very-high computational load due to iterative numerical solutions of the wave equation, as required by industry-standard tools such as Full Waveform Inversion (FWI). For example, in an area with surface dimensions of 4.5 km × 4.5 km, hundreds of seismic shot-gather cubes are required for 3D model reconstruction, leading to Terabytes of recorded data. This paper presents a deep learning solution for the reconstruction of realistic 3D models in the presence of field noise recorded in seismic surveys. We implement and analyze a convolutional encoder-decoder architecture that efficiently processes the entire collection of hundreds of seismic shot-gather cubes. The proposed solution demonstrates that realistic 3D models can be reconstructed with a structural similarity index measure (SSIM) of 0.9143 (out of 1.0) in the presence of field noise at 10 dB signal-to-noise ratio.