Meniscal lesion detection and characterization in adult knee MRI: A deep learning model approach with external validation

PurposeEvaluation of a deep learning approach for the detection of meniscal tears and their characterization (presence/absence of migrated meniscal fragment).MethodsA large annotated adult knee MRI database was built combining medical expertise of radiologists and data scientists’ tools. Coronal and sagittal proton density fat suppressed-weighted images of 11,353 knee MRI examinations (10,401 individual patients) paired with their standardized structured reports were retrospectively collected. After database curation, deep learning models were trained and validated on a subset of 8058 examinations. Algorithm performance was evaluated on a test set of 299 examinations reviewed by 5 musculoskeletal specialists and compared to general radiologists’ reports. External validation was performed using the publicly available MRNet database. Receiver Operating Characteristic (ROC) curves results and Area Under the Curve (AUC) values were obtained on internal and external databases.ResultsA combined architecture of meniscal localization and lesion classification 3D convolutional neural networks reached AUC values of 0.93 (95% CI 0.82, 0.95) for medial and 0.84 (95% CI 0.78, 0.89) for lateral meniscal tear detection, and 0.91 (95% CI 0.87, 0.94) for medial and 0.95 (95% CI 0.92, 0.97) for lateral meniscal tear migration detection. External validation of the combined medial and lateral meniscal tear detection models resulted in an AUC of 0.83 (95% CI 0.75, 0.90) without further training and 0.89 (95% CI 0.82, 0.95) with fine tuning.ConclusionOur deep learning algorithm demonstrated high performance in knee menisci lesion detection and characterization, validated on an external database.     

The confluence of machine learning and multiscale simulations

Multiscale modeling has a long history of use in structural biology, as computational biologists strive to overcome the time- and length-scale limits of atomistic molecular dynamics. Contemporary machine learning techniques, such as deep learning, have promoted advances in virtually every field of science and engineering and are revitalizing the traditional notions of multiscale modeling. Deep learning has found success in various approaches for distilling information from fine-scale models, such as building surrogate models and guiding the development of coarse-grained potentials. However, perhaps its most powerful use in multiscale modeling is in defining latent spaces that enable efficient exploration of conformational space. This confluence of machine learning and multiscale simulation with modern high-performance computing promises a new era of discovery and innovation in structural biology.     

Biomechanical modeling of deep squatting: Effects of the interface contact between posterior thigh and shank

Epidemiological studies indicate that occupational activities that require extended deep knee flexion or kneeling are associated with a higher prevalence of knee osteoarthritis. In many sport activities, such as a catcher in a baseball or a softball game, athletes have to make repetitive deep squatting motions, which have been associated with the development of osteochondritis dissecans. Excessive deep knee flexion postures may cause excessive loading in the knee joint. In deep knee flexion postures, the posterior aspect of the shank will contact the posterior thigh, resulting in a compressive force within the soft tissues. The current study was aimed at analyzing the effects of the posterior thigh/shank contact on the joint loading during deep knee flexion in a natural knee. An existing, whole body model with detailed anatomical components of the knee (AnyBody) has been adopted and modified for this study. The effects of the posterior thigh/shank contact were evaluated by comparing the results of the inverse dynamic analysis for two scenarios: with and without the posterior thigh/shank contact force. Our results showed that, in a deep squatting posture (knee flexion 120+ degrees), the posterior thigh/shank contact helps reduce the patellofemoral (PF) and tibiofemoral (TF) normal contact forces by 42% and 57%, respectively.     

Improving protein tertiary structure prediction by deep learning and distance prediction in CASP14

Substantial progresses in protein structure prediction have been made by utilizing deep-learning and residue-residue distance prediction since CASP13. Inspired by the advances, we improve our CASP14 MULTICOM protein structure prediction system by incorporating three new components: (a) a new deep learning-based protein inter-residue distance predictor to improve template-free (ab initio) tertiary structure prediction, (b) an enhanced template-based tertiary structure prediction method, and (c) distance-based model quality assessment methods empowered by deep learning. In the 2020 CASP14 experiment, MULTICOM predictor was ranked seventh out of 146 predictors in tertiary structure prediction and ranked third out of 136 predictors in inter-domain structure prediction. The results demonstrate that the template-free modeling based on deep learning and residue-residue distance prediction can predict the correct topology for almost all template-based modeling targets and a majority of hard targets (template-free targets or targets whose templates cannot be recognized), which is a significant improvement over the CASP13 MULTICOM predictor. Moreover, the template-free modeling performs better than the template-based modeling on not only hard targets but also the targets that have homologous templates. The performance of the template-free modeling largely depends on the accuracy of distance prediction closely related to the quality of multiple sequence alignments. The structural model quality assessment works well on targets for which enough good models can be predicted, but it may perform poorly when only a few good models are predicted for a hard target and the distribution of model quality scores is highly skewed. MULTICOM is available at https://github.com/jianlin-cheng/MULTICOM_Human_CASP14/tree/CASP14_DeepRank3 and https://github.com/multicom-toolbox/multicom/tree/multicom_v2.0 .     

Predicting human chronically paralyzed muscle force: a comparison of three mathematical models.

Chronic spinal cord injury (SCI) induces detrimental musculoskeletal adaptations that adversely affect health status, ranging from muscle paralysis and skin ulcerations to osteoporosis. SCI rehabilitative efforts may increasingly focus on preserving the integrity of paralyzed extremities to maximize health quality using electrical stimulation for isometric training and/or functional activities. Subject-specific mathematical muscle models could prove valuable for predicting the forces necessary to achieve therapeutic loading conditions in individuals with paralyzed limbs. Although numerous muscle models are available, three modeling approaches were chosen that can accommodate a variety of stimulation input patterns. To our knowledge, no direct comparisons between models using paralyzed muscle have been reported. The three models include 1) a simple second-order linear model with three parameters and 2) two six-parameter nonlinear models (a second-order nonlinear model and a Hill-derived nonlinear model). Soleus muscle forces from four individuals with complete, chronic SCI were used to optimize each model's parameters (using an increasing and decreasing frequency ramp) and to assess the models' predictive accuracies for constant and variable (doublet) stimulation trains at 5, 10, and 20 Hz in each individual. Despite the large differences in modeling approaches, the mean predicted force errors differed only moderately (8-15% error; P=0.0042), suggesting physiological force can be adequately represented by multiple mathematical constructs. The two nonlinear models predicted specific force characteristics better than the linear model in nearly all stimulation conditions, with minimal differences between the two nonlinear models. Either nonlinear mathematical model can provide reasonable force estimates; individual application needs may dictate the preferred modeling strategy.     

Effects of a knee extension constraint brace on selected lower extremity motion patterns during a stop-jump task

Small knee flexion angle during landing has been proposed as a potential risk factor for sustaining noncontact ACL injury. A brace that promotes increased knee flexion and decreased posterior ground reaction force during landing may prove to be advantageous for developing prevention strategies. Forty male and forty female recreational athletes were recruited. Three-dimensional videographic and ground reaction force data in a stop-jump task were collected in three conditions. Knee flexion angle at peak posterior ground reaction force, peak posterior ground reaction force, the horizontal velocity of approach run, the vertical velocity at takeoff, and the knee flexion angle at takeoff were compared among conditions: knee extension constraint brace, nonconstraint brace, and no brace. The knee extension constraint brace significantly increased knee flexion angle at peak posterior ground reaction force. Both knee extension constraint brace and nonconstraint brace significantly decreased peak posterior ground reaction force during landing. The brace and knee extension constraint did not significantly affect the horizontal velocity of approach run, the vertical velocity at takeoff, and the knee flexion angle at takeoff. A knee extension constraint brace exhibits the ability to modify the knee flexion angle at peak posterior ground reaction force and peak posterior ground reaction force during landing.     

The interaction of muscle moment arm,knee laxity,and torque in a multi-scale musculoskeletal model of the lower limb

IntroductionMusculoskeletal modeling allows insight into the interaction of muscle force and knee joint kinematics that cannot be measured in the laboratory. However, musculoskeletal models of the lower extremity commonly use simplified representations of the knee that may limit analyses of the interaction between muscle forces and joint kinematics. The goal of this research was to demonstrate how muscle forces alter knee kinematics and consequently muscle moment arms and joint torque in a musculoskeletal model of the lower limb that includes a deformable representation of the knee.MethodsTwo musculoskeletal models of the lower limb including specimen-specific articular geometries and ligament deformability at the knee were built in a finite element framework and calibrated to match mean isometric torque data collected from 12 healthy subjects. Muscle moment arms were compared between simulations of passive knee flexion and maximum isometric knee extension and flexion. In addition, isometric torque results were compared with predictions using simplified knee models in which the deformability of the knee was removed and the kinematics at the joint were prescribed for all degrees of freedom.ResultsPeak isometric torque estimated with a deformable knee representation occurred between 45° and 60° in extension, and 45° in flexion. The maximum isometric flexion torques generated by the models with deformable ligaments were 14.6% and 17.9% larger than those generated by the models with prescribed kinematics; by contrast, the maximum isometric extension torques generated by the models were similar. The change in hamstrings moment arms during isometric flexion was greater than that of the quadriceps during isometric extension (a mean RMS difference of 9.8 mm compared to 2.9 mm, respectively).DiscussionThe large changes in the moment arms of the hamstrings, when activated in a model with deformable ligaments, resulted in changes to flexion torque. When simulating human motion, the inclusion of a deformable joint in a multi-scale musculoskeletal finite element model of the lower limb may preserve the realistic interaction of muscle force with knee kinematics and torque.     

Measurement of in vivo anterior cruciate ligament strain during dynamic jump landing

Despite recent attention in the literature, anterior cruciate ligament (ACL) injury mechanisms are controversial and incidence rates remain high. One explanation is limited data on in vivo ACL strain during high-risk, dynamic movements. The objective of this study was to quantify ACL strain during jump landing. Marker-based motion analysis techniques were integrated with fluoroscopic and magnetic resonance (MR) imaging techniques to measure dynamic ACL strain non-invasively. First, eight subjects' knees were imaged using MR. From these images, the cortical bone and ACL attachment sites of the tibia and femur were outlined to create 3D models. Subjects underwent motion analysis while jump landing using reflective markers placed directly on the skin around the knee. Next, biplanar fluoroscopic images were taken with the markers in place so that the relative positions of each marker to the underlying bone could be quantified. Numerical optimization allowed jumping kinematics to be superimposed on the knee model, thus reproducing the dynamic in vivo joint motion. ACL length, knee flexion, and ground reaction force were measured. During jump landing, average ACL strain peaked 55±14 ms (mean and 95% confidence interval) prior to ground impact, when knee flexion angles were lowest. The peak ACL strain, measured relative to its length during MR imaging, was 12±7%. The observed trends were consistent with previously described neuromuscular patterns. Unrestricted by field of view or low sampling rate, this novel approach provides a means to measure kinematic patterns that elevate ACL strains and that provide new insights into ACL injury mechanisms.     

Challenges in structural modeling of RNA-protein interactions

In the past few years, the number of RNA-binding proteins (RBP) and RNA-RBP interactions has increased significantly. Here, we review recent developments in the methodology for protein-RNA and protein–protein complex structure modeling with deep learning and co-evolution, as well as discuss the challenges and opportunities for building a reliable approach for protein-RNA complex structure modelling. Protein Data bank (PDB) and Cross-linking immunoprecipitation (CLIP) data could be combined together and used to infer 2D geometry of protein-RNA interactions by deep learning.     

Using force sensing insoles to predict kinetic knee symmetry during a stop jump

Knee kinetic asymmetries are present during jump-landings in athletes returning to sport following anterior cruciate ligament (ACL) reconstruction, and are associated with an increased risk for sustaining a second ACL injury. The loadsol® is a wireless load sensing insole that can be used in non-laboratory settings. The purpose of this study was to determine if the loadsol® could be used to predict knee extension moment and power symmetry during a bilateral stop jump task in healthy recreational athletes. Forty-two uninjured recreational athletes completed seven bilateral stop jumps. During each landing, the loadsol® (100 Hz) measured plantar load while 3D ground reaction forces (1920 Hz) and lower extremity kinematics (240 Hz) were collected simultaneously. Peak impact force, loading rate, and impulse were quantified using the loadsol® and peak knee extension moment, average knee extension moment, and total knee work was quantified using the laboratory instrumentation. Limb symmetry indices were quantified for each outcome measure. Multivariate backwards regressions were used to determine if loadsol® symmetry could predict knee kinetic symmetry. Intraclass correlation coefficients (ICCs) and Bland-Altman plots were used to determine the agreement and error between predicted and actual knee kinetic symmetry. Loadsol® impulse and peak impact force symmetry significantly predicted kinetic knee symmetry and explained 42–61% of its variance. There was good agreement (ICCs = 0.742–0.862) between predicted and actual knee kinetic symmetry, and the error in the predicted outcomes range from ±18 to ±43. These results support using the loadsol® to screen for kinetic symmetries during landing in athletes following ACL reconstruction.     

Protein tertiary structure modeling driven by deep learning and contact distance prediction in CASP13

Predicting residue-residue distance relationships (eg, contacts) has become the key direction to advance protein structure prediction since 2014 CASP11 experiment, while deep learning has revolutionized the technology for contact and distance distribution prediction since its debut in 2012 CASP10 experiment. During 2018 CASP13 experiment, we enhanced our MULTICOM protein structure prediction system with three major components: contact distance prediction based on deep convolutional neural networks, distance-driven template-free (ab initio) modeling, and protein model ranking empowered by deep learning and contact prediction. Our experiment demonstrates that contact distance prediction and deep learning methods are the key reasons that MULTICOM was ranked 3rd out of all 98 predictors in both template-free and template-based structure modeling in CASP13. Deep convolutional neural network can utilize global information in pairwise residue-residue features such as coevolution scores to substantially improve contact distance prediction, which played a decisive role in correctly folding some free modeling and hard template-based modeling targets. Deep learning also successfully integrated one-dimensional structural features, two-dimensional contact information, and three-dimensional structural quality scores to improve protein model quality assessment, where the contact prediction was demonstrated to consistently enhance ranking of protein models for the first time. The success of MULTICOM system clearly shows that protein contact distance prediction and model selection driven by deep learning holds the key of solving protein structure prediction problem. However, there are still challenges in accurately predicting protein contact distance when there are few homologous sequences, folding proteins from noisy contact distances, and ranking models of hard targets.     

Deep template-based protein structure prediction

MotivationProtein structure prediction has been greatly improved by deep learning, but most efforts are devoted to template-free modeling. But very few deep learning methods are developed for TBM (template-based modeling), a popular technique for protein structure prediction. TBM has been studied extensively in the past, but its accuracy is not satisfactory when highly similar templates are not available.ResultsThis paper presents a new method NDThreader (New Deep-learning Threader) to address the challenges of TBM. NDThreader first employs DRNF (deep convolutional residual neural fields), which is an integration of deep ResNet (convolutional residue neural networks) and CRF (conditional random fields), to align a query protein to templates without using any distance information. Then NDThreader uses ADMM (alternating direction method of multipliers) and DRNF to further improve sequence-template alignments by making use of predicted distance potential. Finally, NDThreader builds 3D models from a sequence-template alignment by feeding it and sequence coevolution information into a deep ResNet to predict inter-atom distance distribution, which is then fed into PyRosetta for 3D model construction. Our experimental results show that NDThreader greatly outperforms existing methods such as CNFpred, HHpred, DeepThreader and CEthreader. NDThreader was blindly tested in CASP14 as a part of RaptorX server, which obtained the best average GDT score among all CASP14 servers on the 58 TBM targets.     

Analysis of distance-based protein structure prediction by deep learning in CASP13

This paper reports the CASP13 results of distance-based contact prediction, threading, and folding methods implemented in three RaptorX servers, which are built upon the powerful deep convolutional residual neural network (ResNet) method initiated by us for contact prediction in CASP12. On the 32 CASP13 FM (free-modeling) targets with a median multiple sequence alignment (MSA) depth of 36, RaptorX yielded the best contact prediction among 46 groups and almost the best 3D structure modeling among all server groups without time-consuming conformation sampling. In particular, RaptorX achieved top L/5, L/2, and L long-range contact precision of 70%, 58%, and 45%, respectively, and predicted correct folds (TMscore > 0.5) for 18 of 32 targets. Further, RaptorX predicted correct folds for all FM targets with >300 residues (T0950-D1, T0969-D1, and T1000-D2) and generated the best 3D models for T0950-D1 and T0969-D1 among all groups. This CASP13 test confirms our previous findings: (a) predicted distance is more useful than contacts for both template-based and free modeling; and (b) structure modeling may be improved by integrating template and coevolutionary information via deep learning. This paper will discuss progress we have made since CASP12, the strength and weakness of our methods, and why deep learning performed much better in CASP13.     

An integrated approach based on the correction of imbalanced small datasets and the application of machine learning algorithms to predict total phosphorus concentration in rivers

Increased concentrations of Total Phosphorus (TP) in freshwater systems lead to eutrophication and can contribute to a wide range of environmental effects. In the modern era, water quality models have increasingly been used globally for the development of management scenarios with the aim of reducing the eutrophication risk. However, the accuracy of these models is limited by the quality of the boundary conditions forcing data, namely TP concentration datasets. In this study, a novel methodology is proposed to improve machine learning prediction accuracy in the modeling of river TP concentration forced with small input training datasets. These models can then be used to increase the quality and consistency of the TP concentration datasets required to force water quality models. This new methodology relies on the generation of 100 new training datasets from the raw training datasets of input predictors through the implementation of an over/undersampling technique. The modeling approach used in this study was supported by the application of ten machine learning algorithms to estimate the TP concentration values in 22 rivers located in Portugal. The modeling approach also included an input feature importance evaluation, as well as model hyperparameter optimization. In general terms, the Extreme Gradient Boosting (XGBoost) and Support Vector Regressor (SVR) models performed best overall, with the ensemble results recorded for both models working to increase the mean Nash-Sutcliffe efficiency (NSE) across all the areas being studied by 96% (0.01 ± 0.22 to 0.31 ± 0.32) and reduce the mean percentage bias (PBIAS) by 43% (18.47 ± 17.31 to 10.60 ± 17.40). The results of this study suggest that the solution proposed has the potential to significantly improve the modeling of TP concentration in rivers with machine learning methods, as well as providing increased scope for its application to larger training datasets and the prediction of other types of dependent variables. Hopefully, the results of this study will further add to the body of information available in this area of research and aid the development of the water management process.     

An active learning approach to investigate the ecosystem of tide flats using 3D modeling and printing

ABSTRACT

This paper presents an active learning approach that focuses on practical investigation of the ecosystem of tidal flats using 3D modeling and printing for biology students in order to enhance understanding of natural selection. The learning approach for the study followed a 5-step procedure: i) learning about 3D modeling and printing, ii) exploration of the ecosystem of tidal flats, iii) 3D designing of a bird beak, iv) 3D printing of a constructed beak, and v) a natural selection simulation activity. The learning method presented in this study centered on active student exploration of the tidal flats ecosystem using 3D modeling and printing. The learning approach presented in this paper could be implicated at schools to aid in students’ understanding of natural selection as it allows students to firsthand examine simulation changes to a bird beak and benthic communities. This study suggested the active learning method for natural selection as it incorporates student-designed exploration and direct investigative appraisal of the selection process.     

Verification of predicted specimen-specific natural and implanted patellofemoral kinematics during simulated deep knee bend

Verified computational models represent an efficient method for studying the relationship between articular geometry, soft-tissue constraint, and patellofemoral (PF) mechanics. The current study was performed to evaluate an explicit finite element (FE) modeling approach for predicting PF kinematics in the natural and implanted knee. Experimental three-dimensional kinematic data were collected on four healthy cadaver specimens in their natural state and after total knee replacement in the Kansas knee simulator during a simulated deep knee bend activity. Specimen-specific FE models were created from medical images and CAD implant geometry, and included soft-tissue structures representing medial–lateral PF ligaments and the quadriceps tendon. Measured quadriceps loads and prescribed tibiofemoral kinematics were used to predict dynamic kinematics of an isolated PF joint between 10° and 110° femoral flexion. Model sensitivity analyses were performed to determine the effect of rigid or deformable patellar representations and perturbed PF ligament mechanical properties (pre-tension and stiffness) on model predictions and computational efficiency.Predicted PF kinematics from the deformable analyses showed average root mean square (RMS) differences for the natural and implanted states of less than 3.1° and 1.7 mm for all rotations and translations. Kinematic predictions with rigid bodies increased average RMS values slightly to 3.7° and 1.9 mm with a five-fold decrease in computational time. Two-fold increases and decreases in PF ligament initial strain and linear stiffness were found to most adversely affect kinematic predictions for flexion, internal–external tilt and inferior–superior translation in both natural and implanted states. The verified models could be used to further investigate the effects of component alignment or soft-tissue variability on natural and implant PF mechanics.     

Modeling cellular processes in 3D

Recent advances in photonic imaging and fluorescent protein technology offer unprecedented views of molecular space-time dynamics in living cells. At the same time, advances in computing hardware and software enable modeling of ever more complex systems, from global climate to cell division. As modeling and experiment become more closely integrated we must address the issue of modeling cellular processes in 3D. Here, we highlight recent advances related to 3D modeling in cell biology. While some processes require full 3D analysis, we suggest that others are more naturally described in 2D or 1D. Keeping the dimensionality as low as possible reduces computational time and makes models more intuitively comprehensible; however, the ability to test full 3D models will build greater confidence in models generally and remains an important emerging area of cell biological modeling.     

Leaving flatland: Advances in 3D behavioral measurement

Animals move in three dimensions (3D). Thus, 3D measurement is necessary to report the true kinematics of animal movement. Existing 3D measurement techniques draw on specialized hardware, such as motion capture or depth cameras, as well as deep multi-view and monocular computer vision. Continued advances at the intersection of deep learning and computer vision will facilitate 3D tracking across more anatomical features, with less training data, in additional species, and within more natural, occlusive environments. 3D behavioral measurement enables unique applications in phenotyping, investigating the neural basis of behavior, and designing artificial agents capable of imitating animal behavior.     

Biomechanical but not timed performance asymmetries persist between limbs 9 months after ACL reconstruction during planned and unplanned change of direction

Whilst anterior cruciate ligament injury commonly occurs during change of direction (CoD) tasks, there is little research on how athletes execute CoD after anterior cruciate ligament reconstruction (ACLR). The aims of this study were to determine between-limb and between-test differences in performance (time) and joint kinematics and kinetics during planned and unplanned CoD. One hundred and fifty-six male subjects carried out 90° maximal effort, planned and unplanned CoD tests in a 3D motion capture laboratory 9 months after ACLR. Statistical parametric mapping (2 × 2 ANOVA; limb × test) was used to identify differences in CoD time and biomechanical measures between limbs and between tests. There was no interaction effect but a main effect for limb and task. There was no between-limb difference in the time to complete both CoD tests. Between-limb differences were found for internal knee valgus moment, knee internal rotation and flexion angle, knee extension and external rotation moment and ankle external rotation moment with lower values on the ACLR side (effect size 0.72–0.5). Between test differences were found with less contralateral pelvis rotation, distance from centre of mass to the ankle in frontal plane, posterior ground reaction force and greater hip abduction during the unplanned CoD (effect size 0.75–0.5). Findings demonstrated that kinematic and kinetic differences between limbs are evident during both CoD tests 9 months after surgery, despite no statistical differences in performance time. Biomechanical differences between tests were found in variables, which have previously been associated with ACL injury mechanism during unplanned CoD.