Publications from the lab.

. Deep learning models have demonstrated remarkable success in fields such as language processing and computer vision, routinely employed for tasks like language translation, image classification, and anomaly detection. Recent advancements in ocean sciences, particularly in data assimilation (DA), suggest that machine learning can emulate dynamical models, replace traditional DA steps to expedite processes, or serve as hybrid surrogate models to enhance forecasts. However, these studies often rely on ocean models of intermediate complexity, which involve significant simplifications that present challenges when transitioning to full-scale operational ocean models. This work explores the application of convolutional neural networks (CNNs) in data assimilation within the context of the HYbrid Coordinate Ocean Model (HYCOM) in the Gulf of Mexico. The CNNs are trained to correct model errors from a 2-year, highresolution (1/25°) HYCOM dataset, assimilated using the Tendral Statistical Interpolation System (T-SIS). The CNNs are trained to replicate the increments generated by the TSIS data assimilation package, aiming to correct model forecasts of sea surface temperature (SST) and sea surface height (SSH). The inputs to the CNNs include real satellite observations of SST from the Group for High Resolution Sea Surface Temperature (GHRSST), along-track altimeter SSH observations (ADT), the model background state (previous forecast), and the innovations (differences between observations and background). We assess the performance of the CNNs across five controlled experiments, designed to provide insights into their application in environments governed by full primitive equations, real observations, and complex topographies. The experiments focus on evaluating (1) the architecture and complexity of the CNNs, (2) the type and quantity of observations, (3) the type and number of assimilated fields, (4) the impact of training window size, and (5) the influence of coastal boundaries. Our findings reveal significant correlations between the chosen training window size – a factor not commonly examined – and the CNNs’ ability to assimilate observations effectively. We also establish a clear link between the CNNs’ architecture and complexity and their overall performance. This research uses artificial intelligence to enhance ocean forecasting in the Gulf of Mexico. By using convolutional neural networks, the study improves predictions of sea temperatures and heights by integrating real satellite data with existing models. Through five comprehensive experiments, the team found that the amount of training data and the design of the neural networks significantly affect accuracy. These insights pave the way for faster, more reliable ocean models, benefiting environmental monitoring and maritime operations. 1

Lagrangian coherent eddies efficiently transport water properties, such as heat and salt, as well as tracers, including oil, larvae, and Sargassum, throughout the ocean. For instance, during the 2010 Deepwater Horizon oil spill, part of the oil was captured within a Loop Current Frontal Eddy (LCFE), preventing it from reaching the Florida Keys. Similarly, Loop Current Eddies (LCEs) carry warmer, saltier waters typical of the Caribbean Sea to the western Gulf of Mexico (GoM). In this study, we employ machine learning alongside various satellite observations—absolute dynamic topography (ADT), sea surface temperature (SST), and chlorophyll‐a (Chl‐a)—to identify Lagrangian coherent eddies in the GoM and predict their lifetime. Three durations of Lagrangian coherence are investigated: 5, 10, and 20 days. This study also investigates the contributions of Chl‐a to identifying and forecasting LCEs' and LCFEs' Lagrangian coherence, aiming to assess the advantages of integrating this data set into data‐assimilative Gulf ocean models, in addition to ADT and SST. The machine learning model trained with ADT successfully identifies and predicts the lifetimes of eddies, achieving accuracy rates of 90% for LCE identification and 93% for lifetime prediction, along with 71% and 61% for LCFEs, respectively. Incorporating SST and Chl‐a enhanced eddy predictions over ADT‐only or ADT and SST combined, in particular LCEs and LCFEs, highlighting the benefits of assimilating Chl‐a into ocean models to improve the representation and the forecast of these eddies. This machine learning framework has the potential to advance predictions of eddy lifetimes and the advection of various tracers. Plain Language Summary Lagrangian coherent eddies are types of vortices in the ocean that trap water in their interior and transport it without exchange with the exterior water. These eddies play a key role in transporting water properties such as heat and salt, as well as tracers such as oil, larvae, and seaweed (e.g., Sargassum) across the ocean. For example, during the 2010 Deepwater Horizon oil spill, a type of eddy called a Loop Current Frontal Eddy (LCFE) trapped some of the oil, keeping it from reaching the Florida Keys. This study uses machine learning and satellite data—sea surface height, sea surface temperature, and chlorophyll concentration—to identify and predict the lifetimes of Lagrangian coherent eddies in the GoM. Three durations of eddy coherence (5, 10, and 20 days) are analyzed. The machine learning model trained with sea surface height successfully identifies and predicts the lifetimes of eddies, achieving accuracy rates of 90% for Loop Current Eddies identification and 93% for lifetime prediction, along with 71% and 61% for LCFEs, respectively. Adding chlorophyll data from satellites improved the predictions compared to using sea surface height and temperature alone. This machine learning framework can advance predictions of eddy lifetimes and tracer transport.

Accurate circulation modeling in the Gulf of Mexico (GoM) is hampered by the limited availability of insitu subsurface data, leading to inaccuracies in subsurface representations. These inaccuracies diminish the reliability of ocean models and constrain the duration of dependable forecasts. This study introduces NeSPReSO (Neural Synthetic Profiles from Remote Sensing and Observations), a data-driven method to efficiently and accurately estimate subsurface temperature and salinity profiles using satellite-derived surface data. This provides an alternative to conventional synthetic data generation techniques. Principal Component Analysis (PCA) is applied to extract the main features of temperature and salinity profiles of an Argo dataset. Then, a neural network is trained to predict these principal features using inputs such as time, location, and satellite-derived absolute dynamic topography alongside sea surface temperature and salinity. The model, evaluated using additional Argo profiles and glider data collected in the Gulf of Mexico, over-performs other traditional synthetic data generation methods, such as the Gravest Empirical Modes (GEM), Multiple Linear Regression (MLR) and Improved Synthetic Ocean Profile (ISOP), in terms of root mean square error and bias. Our findings indicate that our method effectively captures the main variations of subsurface fields, and that synthetic profiles generated by the model align well with actual observations, accurately capturing key features such as thermoclines, haloclines, and temperature-salinity structure of the region. This new method will be implemented in GoM data assimilative models and is expected to improve the accuracy of modeled subsurface currents.

Mexico City, a densely populated urban area, experiences multiple episodes of elevated air pollution almost every year. To mitigate the impact of these pollution episodes on the population, it is important to improve forecast systems that allow government authorities to take preventive actions, reducing the exposure of vulnerable groups. This study introduces a pilot operational ozone forecasting system based on machine learning. The proposed system is trained using historical data from a long-standing governmental air quality and atmospheric monitoring network, and with meteorological reanalysis data from a regional implementation of the Weather Research and Forecasting (WRF) model for Mexico. Additional input features are incorporated, including cyclical time encoding for the day of the week, time of day, and day of the year, to improve system accuracy. Once trained, the system utilizes real-time data from multiple atmospheric stations and regional meteorological forecasts to predict ozone levels for the following 24 h. This study evaluates the impact of data augmentation and the advantages of integrating meteorological forecast information into the model. The model achieves a mean absolute error (MAE) of 9.81 ppb and an index of agreement of 0.91 across all stations. For the top 10 stations, the MAE falls below 8.7 ppb, and the index of agreement exceeds 0.93. The system’s performance is comparable to similar systems in other large metropolitan areas and represents an improvement over the existing systems in Mexico City. This operational system is available at https://aire.atmosfera.unam.mx/.

Uterine muscle contractility is essential for reproductive processes including sperm and embryo transport, and during the uterine cycle to remove menstrual effluent. Even still, uterine contractions have primarily been studied in the context of preterm labor. This is partly due to a lack of methods for studying the uterine muscle contractility in the intact organ. Here, we describe an imaging-based method to evaluate mouse uterine contractility of both the longitudinal and circular muscles in the cycling stages and in early pregnancy. By transforming the imagebased data into three-dimensional spatiotemporal contractility maps, we calculate waveform characteristics of muscle contractions, including amplitude, frequency, wavelength, and velocity. We report that the native organ is highly contractile during the progesterone-dominant diestrus stage of the cycle when compared to the estrogen-dominant proestrus and estrus stages. We also observed that during the first phase of uterine embryo movement when clustered embryos move toward the middle of the uterine horn, contractions are dynamic and non-uniform between different segments of the uterine horn. In the second phase of embryo movement, contractions are more uniform and rhythmic throughout the uterine horn. Finally, in Lpar3 −/− uteri, which display faster embryo movement, we observe global and regional increases in contractility. Our method provides a means to understand the wave characteristics of uterine smooth muscle in response to modulators and in genetic mutants. Better understanding uterine contractility in the early pregnancy stages is critical for the advancement of artificial reproductive technologies and a possibility of modulating embryo movement during clinical embryo transfers. Summary Sentence The imaging-based method of ex vivo uterine contractility demonstrates quantitative differences in contractions during estrous cycle, preimplantation stages of pregnancy, and in LPAR3-deficient mutants. Received: December 3, 2023. Revised: April 3, 2024. Accepted: May 2, 2024 © The Author(s) 2024. Published by Oxford University Press on behalf of Society for the Study of Reproduction. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. Downloaded from https://academic.oup.com/biolreprod/article/110/6/1175/7666628 by guest on 18 November 2024 1 Department of Obstetrics, Gynecology and Reproductive Biology, Michigan State University, East Lansing, Michigan, USA 1176 Imaging murine uterine contractions, 2024, Vol. 110, No. 6 Graphical Abstract

The persistent increase in marine plastic litter has become a major global concern, with one of the highest plastic concentrations in the world’s oceans found in the Wider Caribbean Region (WCR). In this study, we use marine plastic litter tracking simulations to investigate where marine plastic accumulates, i.e., hotspots, in the WCR and how the accumulation varies on seasonal timescales. We show that most of the marine plastic waste converges on the coastlines shortly after being released into the WCR because of the strong surface current and the predominant easterly winds. Major plastic accumulations take place along (i) the western coastline of the WCR, especially the north–south-oriented coasts of Costa Rica/Nicaragua, Guatemala/Belize/Mexico, and Texas, and (ii) the coastlines of Haiti–Dominican Republic and Venezuela. Relatively low plastic accumulation is found along western Florida, the western Yucatán peninsula, and the leeward and windward Caribbean islands. Accumulation along the western WCR coastlines is modulated primarily by ocean currents and exhibits significant seasonal variabilities due to changes in wind patterns. The accumulation observed on the Haiti–Dominican Republic and Venezuela coastlines is primarily due to the proximity of large, mismanaged plastic waste sources. Finally, we discuss the uncertainty associated with the choices made in defining the different criteria for plastic beaching in the models. Citation: Xu, X.; Chassignet, E.P.;

This study utilizes Deep Neural Networks (DNN) to improve the K‐Profile Parameterization Supporting Information: Supporting Information may be found in the online version of this article. Correspondence to: J.‐H. Liang, jliang@lsu.edu Citation: Yuan, J., Liang, J.‐H., Chassignet, E. P., Zavala‐Romero, O., Wan, X., & Cronin, M. F. (2024). The K‐profile parameterization augmented by deep neural networks (KPP_DNN) in the general ocean turbulence model (GOTM). Journal of Advances in Modeling Earth Systems, 16, e2024MS004405. https://doi. org/10.1029/2024MS004405 Received 19 APR 2024 Accepted 4 SEP 2024 © 2024 The Author(s). Journal of Advances in Modeling Earth Systems published by Wiley Periodicals LLC on behalf of American Geophysical Union. This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non‐commercial and no modifications or adaptations are made. YUAN ET AL. (KPP) for the vertical mixing effects in the ocean's surface boundary layer turbulence. The deep neural networks were trained using 11‐year turbulence‐resolving solutions, obtained by running a large eddy simulation model for Ocean Station Papa, to predict the turbulence velocity scale coefficient and unresolved shear coefficient in the KPP. The DNN‐augmented KPP schemes (KPP_DNN) have been implemented in the General Ocean Turbulence Model (GOTM). The KPP_DNN is stable for long‐term integration and more efficient than existing variants of KPP schemes with wave effects. Three different KPP_DNN schemes, each differing in their input and output variables, have been developed and trained. The performance of models utilizing the KPP_DNN schemes is compared to those employing traditional deterministic first‐order and second‐moment closure turbulent mixing parameterizations. Solution comparisons indicate that the simulated mixed layer becomes cooler and deeper when wave effects are included in parameterizations, aligning closer with observations. In the KPP framework, the velocity scale of unresolved shear, which is used to calculate ocean surface boundary layer depth, has a greater impact on the simulated mixed layer than the magnitude of diffusivity does. In the KPP_DNN, unresolved shear depends not only on wave forcing, but also on the mixed layer depth and buoyancy forcing. Plain Language Summary The uppermost tens of meters of the ocean, known as the ocean surface boundary layer, are rich in intricate and chaotic fine‐scale (cm to 100s m) ocean currents referred to as turbulence. These currents, spanning from centimeters to hundreds of meters, play pivotal roles in shaping the oceanic environment and influencing Earth's climate dynamics. Despite their significance, simulating these fine‐scale ocean currents remains beyond the capabilities of current and foreseeable supercomputing resources. Consequently, simplified formulas derived from fundamental principles are commonly employed to approximate these currents in ocean and climate models. However, these approximations still cannot cover all types of choppy currents and uncertainties in these approximations represent a substantial source of bias in contemporary ocean and climate modeling endeavors. In this study, we enhance one of the prevalent physics‐ based approximations of fine‐scale turbulent currents using machine learning techniques. Our tests show that integrating machine learning in physics‐based approximation is stable and efficient and is suitable for use in ocean and climate models.

During the period of 2018–2022, there were six named Loop Current Eddy (LCE) shedding events in the central Gulf of Mexico (GoM). LCEs form when a large anticyclonic eddy (AE) separates from the main Loop Current (LC) and propagates westward. In doing so, each LCE traps and advects warmer, saltier waters with lower Chlorophyll-a (Chl-a) concentrations than the surrounding Gulf waters. This difference in water mass permits the study of the effectiveness of using Chl-a from satellite-derived ocean color to identify LCEs in the GoM. In this work, we apply an eddy-tracking algorithm to Chl-a to detect LCEs, which we have validated against the traditional sea surface height(SSH) based eddy-tracking approach with three datasets. We apply a closed-contour eddy-tracking algorithm to the SSH of two model products (HYbrid Coordination Ocean Model; HYCOM and Nucleus for European Modelling of the Ocean; NEMO) and absolute dynamic topography (ADT) from altimetry, as well as satellite-derived Chl-a data to identify the six named LCEs from 2018 to 2022. We find that Chl-a best characterizes LCEs in the summertime due to a basin-wide increase in the horizontal gradient of Chl-a, which permits a more clearly defined eddy edge. This study demonstrates that Chl-a can be effectively used to identify and track LC and LCEs in the GoM, serving as a promising source of information for regional data assimilative models. Citation: Trott, C.B.; Subrahmanyam, B.; Hiron, L.; Zavala-Romero, O.

The utilization of multi-parametric MRI (mpMRI) in clinical decisions regarding prostate cancer patients’ management has recently increased. After biopsy, clinicians can assess risk using National Comprehensive Cancer Network (NCCN) risk stratification schema and commercially available genomic classifiers, such as Decipher. We built radiomics-based models to predict lesions/patients at low risk prior to biopsy based on an established three-tier clinical-genomic classification system. Radiomic features were extracted from regions of positive biopsies and Normally Appearing Tissues (NAT) on T2-weighted and Diffusion-weighted Imaging. Using only clinical information available prior to biopsy, five models for predicting low-risk lesions/patients were evaluated, based on: 1: Clinical variables; 2: Lesion-based radiomic features; 3: Lesion and NAT radiomics; 4: Clinical and lesion-based radiomics; and 5: Clinical, lesion and NAT radiomic features. Eighty-three mpMRI exams from 78 men were analyzed. Models 1 and 2 performed similarly (Area under the receiver operating characteristic curve were 0.835 and 0.838, respectively), but radiomics significantly improved the lesion-based performance of the model in a subset analysis of patients with a negative Digital Rectal Exam (DRE). Adding normal tissue radiomics significantly improved the performance in all cases. Similar patterns were observed on patient-level models. To the best of our knowledge, this is the first study to demonstrate that machine learning radiomics-based models can predict patients’ risk using combined clinical-genomic classification. Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article

Computer-aided detection and diagnosis (CAD) systems have the potential to improve robustness and efficiency compared to traditional radiological reading of magnetic resonance imaging (MRI). Fully automated segmentation of the prostate is a crucial step of CAD for prostate cancer, but visual inspection is still required to detect poorly segmented cases. The aim of this work was therefore to establish a fully automated quality control (QC) system for prostate segmentation based on T2-weighted MRI. Four different deep learning-based segmentation methods were used to segment the prostate for 585 patients. First order, shape and textural radiomics features were extracted from the segmented prostate masks. A reference quality score (QS) was calculated for each automated segmentation in comparison to a manual segmentation. A least absolute shrinkage and selection operator (LASSO) was trained and optimized on a randomly assigned training dataset (N = 1756, 439 cases from each segmentation method) to build a generalizable linear regression model based on the radiomics features that best estimated the reference QS. Subsequently, the model was used to estimate the QSs for an independent testing dataset (N = 584, 146 cases from each segmentation method). The mean ± standard deviation absolute error between the estimated and reference QSs was 5.47 ± 6.33 on a scale from 0 to 100. In addition, we found a strong correlation between the estimated and reference QSs (rho = 0.70). In conclusion, we developed an automated QC system that may be helpful for evaluating the quality of automated prostate segmentations.

Purpose Develop a deep-learning-based segmentation algorithm for prostate and its peripheral zone (PZ) that is reliable across multiple MRI vendors. Methods This is a retrospective study. The dataset consisted of 550 MRIs (Siemens-330, General Electric[GE]-220). A multistream 3D convolutional neural network is used for automatic segmentation of the prostate and its PZ using T2-weighted (T2-w) MRI. Prostate and PZ were manually contoured on axial T2-w. The network uses axial, coronal, and sagittal T2-w series as input. The preprocessing of the input data includes bias correction, resampling, and image normalization. A dataset from two MRI vendors (Siemens and GE) is used to test the proposed network. Six different models were trained, three for the prostate and three for the PZ. Of the three, two were trained on data from each vendor separately, and a third (Combined) on the aggregate of the datasets. The Dice coefficient (DSC) is used to compare the manual and predicted segmentation. Results For prostate segmentation, the Combined model obtained DSCs of 0.893 ± 0.036 and 0.825 ± 0.112 (mean ± standard deviation) on Siemens and GE, respectively. For PZ, the best DSCs were from the Combined model: 0.811 ± 0.079 and 0.788 ± 0.093. While the Siemens model underperformed on the GE dataset and vice versa, the Combined model achieved robust performance on both datasets. Conclusion The proposed network has a performance comparable to the interexpert variability for segmenting the prostate and its PZ. Combining images from different MRI vendors on the training of the network is of paramount importance for building a universal model for prostate and PZ segmentation.

. We present a radiomics-based approach developed for the SPIE-AAPM-NCI PROSTATEx challenge. The task was to classify clinically significant prostate cancer in multiparametric (mp) MRI. Data consisted of a “training dataset” (330 suspected lesions from 204 patients) and a “test dataset” (208 lesions/140 patients). All studies included T2-weighted (T2-W), proton density-weighted, dynamic contrast enhanced, and diffusionweighted imaging. Analysis of the images was performed using the MIM imaging platform (MIM Software, Cleveland, Ohio). Prostate and peripheral zone contours were manually outlined on the T2-W images. A workflow for rigid fusion of the aforementioned images to T2-W was created in MIM. The suspicious lesion was outlined using the high b-value image. Intensity and texture features were extracted on four imaging modalities and characterized using nine histogram descriptors: 10%, 25%, 50%, 75%, 90%, mean, standard deviation, kurtosis, and skewness (216 features). Three classification methods were used: classification and regression trees (CART), random forests, and adaptive least absolute shrinkage and selection operator (LASSO). In the held out by the organizers test dataset, the areas under the curve (AUCs) were: 0.82 (random forests), 0.76 (CART), and 0.76 (adaptive LASSO). AUC of 0.82 was the fourth-highest score of 71 entries (32 teams) and the highest for feature-based methods. © 2018 Society of Photo-Optical Instrumentation Engineers (SPIE) [DOI: 10.1117/1.JMI.5.3.034502]

Analyzing soft-tissue structures is particularly challenging due to the lack of homologous landmarks that can be reliably identified across time and specimens. This is particularly true when data are to be collected under field conditions. Here, we present a method that combines photogrammetric techniques and geometric morphometrics methods (GMM) to quantify soft tissues for their subsequent volumetric analysis. We combine previously developed methods for landmark data acquisition and processing with a custom program for volumetric computations. Photogrammetric methods are a particularly powerful tool for field studies as they allow for image acquisition with minimal equipment requirements and for the acquisition of the spatial coordinates of points (anatomical landmarks or others) from these images. For our method, a limited number of homologous landmarks, i.e., points that can be found on any specimen independent of space and time, and further distinctive points, which may vary over time, space and subject, are identified on two-dimensional photographs and their three-dimensional coordinates estimated using photogrammetric methods. The three-dimensional configurations are oriented by the spatial principal components (PCs) of the homologous points. Crucially, this last step orients the configuration such that x and y-information (PC1 and PC2 coordinates) constitute an anatomically-defined plane with the zvalues (PC3 coordinate) in the direction of interest for volume computation. The z-coordinates are then used to estimate the volume of the tissue. We validate our method using a physical, geometric model of known dimensions and physical (wax) models designed to approximate perineal swellings in female macaques. To demonstrate the usefulness and potential of our method, we use it to estimate the volumes of Barbary macaque sexual swellings recorded in the field with video images. By analyzing both the artificial data and real monkey swellings, we validate our method’s accuracy and illustrate its potential for application in important areas of biological research. Citation: Sonnweber RS, Stobbe N, Zavala Romero O, Slice DE, Fieder M, et al. (2013) A New Method for the Analysis of Soft Tissues with Data Acquired under Field Conditions. PLoS ONE 8(6): e67521. doi:10.1371/journal.pone.0067521 Editor: Alistair Robert Evans, Monash University, Australia Received October 15, 2012; Accepted May 21, 2013; Published June 24, 2013 Copyright: ß 2013 Sonnweber et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Funding: While writing the manuscript, Ruth Sonnweber and Nina Stobbe were supported by ERC (European Research Council) Advanced Grant SOMACCA (The Syntax of Mind: A Computational Comparative Apporach) (230604) awarded to W. Tecumseh Fitch. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing Interests: The authors have declared that no competing interests exist. * E-mail: ruth-sophie.sonnweber@univie.ac.at . These authors contributed equally to this work. The perineal swellings of female Barbary macaques (Macaca sylvanus) are an example of a soft tissue structure whose 3D nature would be of interest to describe morphometrically. As with many other females of Old World monkeys, these primates develop distinct anogenital swellings during sexually active phases [2]. Swelling size changes over the female cycle and tumescence and de-tumescence were shown to be related to female sex steroids [3,4,5]. What information these secondary sexual characteristics convey is not yet fully understood. All published studies relating to the morphology of this feature have so far been restricted to analyses of two-dimensional data. We chose this morphological trait as an example for