Search

Background: Segmentation of medical image volumes is a time-consuming manual task. Automatic tools are often tailored toward specific patient cohorts, and it is unclear how they behave in other clinical settings. Purpose: To evaluate the performance of the open-source Multi-Planar U-Net (MPUnet), the validated Knee Imaging Quantification (KIQ) framework, and a state-of-the-art two-dimensional (2D) U-Net architecture on three clinical cohorts without extensive adaptation of the algorithms. Study Type: Retrospective cohort study. Subjects: A total of 253 subjects (146 females, 107 males, ages 57 ± 12 years) from three knee osteoarthritis (OA) studies (Center for Clinical and Basic Research [CCBR], Osteoarthritis Initiative [OAI], and Prevention of OA in Overweight Females [PROOF]) with varying demographics and OA severity (64/37/24/53/2 scans of Kellgren and Lawrence [KL] grades 0–4). Field Strength/Sequence: 0.18 T, 1.0 T/1.5 T, and 3 T sagittal three-dimensional fast-spin echo T1w and dual-echo steady-state sequences. Assessment: All models were fit without tuning to knee magnetic resonance imaging (MRI) scans with manual segmentations from three clinical cohorts. All models were evaluated across KL grades. Statistical Tests: Segmentation performance differences as measured by Dice coefficients were tested with paired, two-sided Wilcoxon signed-rank statistics with significance threshold α = 0.05. Results: The MPUnet performed superior or equal to KIQ and 2D U-Net on all compartments across three cohorts. Mean Dice overlap was significantly higher for MPUnet compared to KIQ and U-Net on CCBR ((Formula presented.) vs. (Formula presented.) and (Formula presented.)), significantly higher than KIQ and U-Net OAI ((Formula presented.) vs. (Formula presented.) and (Formula presented.), and not significantly different from KIQ while significantly higher than 2D U-Net on PROOF ((Formula presented.) vs. (Formula presented.), (Formula presented.), and (Formula presented.). The MPUnet performed significantly better on (Formula presented.) KL grade 3 CCBR scans with (Formula presented.) vs. (Formula presented.) for KIQ and (Formula presented.) for 2D U-Net. Data Conclusion: The MPUnet matched or exceeded the performance of state-of-the-art knee MRI segmentation models across cohorts of variable sequences and patient demographics. The MPUnet required no manual tuning making it both accurate and easy-to-use. Level of Evidence: 3. Technical Efficacy: Stage 2.

To determine the optimal combination of imaging and biochemical biomarkers for use in the prediction of knee osteoarthritis (OA) progression.The present study was a nested case-control trial from the Foundation of the National Institutes of Health OA Biomarkers Consortium that assessed study participants with a Kellgren/Lawrence grade of 1-3 who had complete biomarker data available (n = 539 to 550). Cases were participants' knees that had radiographic and pain progression between 24 and 48 months compared to baseline. Radiographic progression only was assessed in secondary analyses. Biomarkers (baseline and 24-month changes) that had a P value of0.10 in univariate analysis were selected, including quantitative cartilage thickness and volume on magnetic resonance imaging (MRI), semiquantitative MRI markers, bone shape and area, quantitative meniscal volume, radiographic progression (trabecular bone texture [TBT]), and serum and/or urine biochemical markers. Multivariable logistic regression models were built using 3 different stepwise selection methods (complex models versus parsimonious models).Among baseline biomarkers, the number of locations affected by osteophytes (semiquantitative), quantitative central medial femoral and central lateral femoral cartilage thickness, patellar bone shape, and semiquantitative Hoffa-synovitis predicted OA progression in most models (C statistic 0.641-0.671). In most models, 24-month changes in semiquantitative MRI markers (effusion-synovitis, meniscal morphologic changes, and cartilage damage), quantitative central medial femoral cartilage thickness, quantitative medial tibial cartilage volume, quantitative lateral patellofemoral bone area, horizontal TBT (intercept term), and urine N-telopeptide of type I collagen predicted OA progression (C statistic 0.680-0.724). A different combination of imaging and biochemical biomarkers (baseline and 24-month change) predicted radiographic progression only, which had a higher C statistic of 0.716-0.832.The present study highlights the combination of biomarkers with potential prognostic utility in OA disease-modifying trials. Properly qualified, these biomarkers could be used to enrich future trials with participants likely to experience progression of knee OA.

Tensor networks are efficient factorisations of high-dimensional tensors into a network of lower-order tensors. They have been most commonly used to model entanglement in quantum many-body systems and more recently are witnessing increased applications in supervised machine learning. In this work, we formulate image segmentation in a supervised setting with tensor networks. The key idea is to first lift the pixels in image patches to exponentially high-dimensional feature spaces and using a linear decision hyper-plane to classify the input pixels into foreground and background classes. The high-dimensional linear model itself is approximated using the matrix product state (MPS) tensor network. The MPS is weight-shared between the non-overlapping image patches resulting in our strided tensor network model. The performance of the proposed model is evaluated on three 2D- and one 3D- biomedical imaging datasets. The performance of the proposed tensor network segmentation model is compared with relevant baseline methods. In the 2D experiments, the tensor network model yields competitive performance compared to the baseline methods while being more resource efficient., Journal extension of our preliminary conference work "Segmenting two-dimensional structures with strided tensor networks", Selvan et al. 2021, available at arXiv:2102.06900. 24 pages, 12 figures. Accepted to be published at the Journal of Machine Learning for Biomedical Imaging, to be updated at https://www.melba-journal.org/papers/2022:005.html

Roots are vital organs for plants, but the assessment of root traits is difficult, particularly in deep soil layers under natural field conditions. A popular technique to investigate root growth under field or semi-field conditions is the use of minirhizotrons. However, the subsequent manual quantification process is time-consuming and prone to error. We developed a multispectral minirhizotron imaging system and a subsequent image analysis strategy for automated root detection. Five wavelengths in the visible (VIS) and near-infrared (NIR) spectrum are used to enhance living roots by a multivariate grouping of pixels based on differences in reflectance; background noise is suppressed by a vesselness enhancement filter. The system was tested against manual analysis of grid intersections for both spring barley (Hordeum vulgare L.) and perennial ryegrass (Lolium perenne L.) cultivars at two time-points. The images of living roots were captured in wet subsoil conditions with dead roots present from a previous crop. Under the soil conditions used in the study, NIR reflectance (940 nm), provided limited ability to separate between rhizosphere components, compared to reflectance in the violet and blue light spectrum (405 nm and 450 nm). Multivariate image analysis of the spectral data, combined with vesselness enhancement and thresholding allowed for automated detection of living roots. Automated image analysis largely replicated the root intensity found during manual grid intersect analysis of the same images. Although some misclassification occurred, caused by elongated structures of dew and chalkstone with similar reflectance pattern as living root, the system provided similar or in some cases improved detection of genotypic differences in the total root length within each tube. The multispectral imaging system allows for automated detection of living roots in minirhizotron studies. The system requires considerably less time than traditional manual recording using grid intersections. The flexible training strategy used for root segmentation offers hope for the transfer to other rhizosphere components and other soil types of interest.

We thank Roemer and Guermazi for their Letter, “Biochemical cartilage changes based on MRI-defined T2 relaxation times do not equal OA detection” (1). In their comments on our paper (2), the authors (1) raise questions about osteoarthritis (OA) incidence, visual signs of disease on the images, MRI-based scoring systems, and radiomics which we address below. In ref. 2, we propose a fully automated machine learning technique for discovery of imaging biomarkers in the cartilage of asymptomatic individuals for prediction of future symptomatic OA. We overcome a long-standing limitation of machine learning regarding explainability, by enabling visual recognition of classification biomarkers. First, regarding OA incidence, the American College of Rheumatology publication (3) cited by the authors (1) recommends diagnosis based on … [↵][1]1To whom correspondence may be addressed. Email: skundu2{at}jhmi.edu. [1]: #xref-corresp-1-1

The development of new nanomaterials for energy technologies is dependent on understanding the intricate relation between material properties and atomic structure. It is, therefore, crucial to be able to routinely characterise the atomic structure in nanomaterials, and a promising method for this task is Pair Distribution Function (PDF) analysis. The PDF can be obtained through Fourier transformation of x-ray total scattering data, and represents a histogram of all interatomic distances in the sample. Going from the distance information in the PDF to a chemical structure is an unassigned distance geometry problem (uDGP), and solving this is often the bottleneck in nanostructure analysis. In this work, we propose to use a Conditional Variational Autoencoder (CVAE) to automatically solve the uDGP to obtain valid chemical structures from PDFs. We use a simple model system of hypothetical mono-metallic nanoparticles containing up to 100 atoms in the face centered cubic (FCC) structure as a proof of concept. The model is trained to predict the assigned distance matrix (aDM) from a simulated PDF of the structure as the conditional input. We introduce a novel representation of structures by projecting them inside a unit sphere and adding additional anchor points or satellites to help in the reconstruction of the chemical structure. The performance of the CVAE model is compared to a Deterministic Autoencoder (DAE) showing that both models are able to solve the uDGP reasonably well. We further show that the CVAE learns a structured and meaningful latent embedding space which can be used to predict new chemical structures.

Objective Gender is a risk factor in the onset of osteoarthritis (OA). The aim of the study was to investigate gender differences in contact area (CA) and congruity index (CI) in the medial tibiofemoral (MTF) joint in 2 different cohorts, quantified automatically from magnetic resonance imaging (MRI). Design The CA and CI markers were validated on 2 different data sets from Center for Clinical and Basic Research (CCBR) and Osteoarthritis Initiative (OAI). The CCBR cohort consisted of 159 subjects and the OAI subcohort consisted of 1,436 subjects. From the MTF joint, the contact area was located and quantified using Euclidean distance transform. Furthermore, the CI was quantified over the contact area by assessing agreement of the first- and second-order general surface features. Then, the gender differences between CA and CI values were evaluated at different stages of radiographic OA. Results Female CAs were significantly higher than male CAs after normalization, male CIs were significantly higher than female CIs after correcting with age and body mass index ( P < 0.05), consistent across the 2 data sets. For the OAI data set, the gender differences were present at all stages of radiographic OA. Conclusion This study demonstrated the gender differences in CA and CI in MTF joints. The higher normalized CA and lower CI values in female knees may be linked with the increased risk of incidence of radiographic OA in females. These differences may help further understand the gender differences and/or to establish gender specific treatment strategies.

Purpose: To organize a multi-institute knee MRI segmentation challenge for characterizing the semantic and clinical efficacy of automatic segmentation methods relevant for monitoring osteoarthritis progression.Materials and Methods: A dataset partition consisting of three-dimensional knee MRI from 88 retrospective patients at two time points (baseline and 1-year follow-up) with ground truth articular (femoral, tibial, and patellar) cartilage and meniscus segmentations was standardized. Challenge submissions and a majority-vote ensemble were evaluated against ground truth segmentations using Dice score, average symmetric surface distance, volumetric overlap error, and coefficient of variation on a holdout test set. Similarities in automated segmentations were measured using pairwise Dice coefficient correlations. Articular cartilage thickness was computed longitudinally and with scans. Correlation between thickness error and segmentation metrics was measured using the Pearson correlation coefficient. Two empirical upper bounds for ensemble performance were computed using combinations of model outputs that consolidated true positives and true negatives.Results: Six teams (T 1-T 6) submitted entries for the challenge. No differences were observed across any segmentation metrics for any tissues (P = .99) among the four top-performing networks (T 2, T 3, T 4, T 6). Dice coefficient correlations between network pairs were high (> 0.85). Per-scan thickness errors were negligible among networks T 1-T 4 (P = .99), and longitudinal changes showed minimal bias (< 0.03 mm). Low correlations (ρ < 0.41) were observed between segmentation metrics and thickness error. The majority-vote ensemble was comparable to top-performing networks (P = .99). Empirical upper-bound performances were similar for both combinations (P = .99).Conclusion: Diverse networks learned to segment the knee similarly, where high segmentation accuracy did not correlate with cartilage thickness accuracy and voting ensembles did not exceed individual network performance.See also the commentary by Elhalawani and Mak in this issue.Keywords: Cartilage, Knee, MR-Imaging, Segmentation © RSNA, 2020Supplemental material is available for this article.

Many recent medical segmentation systems rely on powerful deep learning models to solve highly specific tasks. To maximize performance, it is standard practice to evaluate numerous pipelines with varying model topologies, optimization parameters, pre- & postprocessing steps, and even model cascades. It is often not clear how the resulting pipeline transfers to different tasks. We propose a simple and thoroughly evaluated deep learning framework for segmentation of arbitrary medical image volumes. The system requires no task-specific information, no human interaction and is based on a fixed model topology and a fixed hyperparameter set, eliminating the process of model selection and its inherent tendency to cause method-level over-fitting. The system is available in open source and does not require deep learning expertise to use. Without task-specific modifications, the system performed better than or similar to highly specialized deep learning methods across 3 separate segmentation tasks. In addition, it ranked 5-th and 6-th in the first and second round of the 2018 Medical Segmentation Decathlon comprising another 10 tasks. The system relies on multi-planar data augmentation which facilitates the application of a single 2D architecture based on the familiar U-Net. Multi-planar training combines the parameter efficiency of a 2D fully convolutional neural network with a systematic train- and test-time augmentation scheme, which allows the 2D model to learn a representation of the 3D image volume that fosters generalization.