Optimal surface segmentation with convex priors in irregularly sampled space

Deep learning network with differentiable dynamic programming for retina OCT surface segmentation

Multiple-surface segmentation in optical coherence tomography (OCT) images is a challenging problem, further complicated by the frequent presence of weak image boundaries. Recently, many deep learning-based methods have been developed for this task and yield remarkable performance. Unfortunately, due to the scarcity of training data in medical imaging, it is challenging for deep learning networks to learn the global structure of the target surfaces, including surface smoothness. To bridge this gap, this study proposes to seamlessly unify a U-Net for feature learning with a constrained differentiable dynamic programming module to achieve end-to-end learning for retina OCT surface segmentation to explicitly enforce surface smoothness. It effectively utilizes the feedback from the downstream model optimization module to guide feature learning, yielding better enforcement of global structures of the target surfaces. Experiments on Duke AMD (age-related macular degeneration) and JHU MS (multiple sclerosis) OCT data sets for retinal layer segmentation demonstrated that the proposed method was able to achieve subvoxel accuracy on both datasets, with the mean absolute surface distance (MASD) errors of 1.88 ± 1.96μm and 2.75 ± 0.94μm, respectively, over all the segmented surfaces.

On the water-fat in-phase assumption for quantitative susceptibility mapping

PURPOSE

To (a) define multi-peak fat model-based effective in-phase echo times for quantitative susceptibility mapping (QSM) in water-fat regions, (b) analyze the relationship between fat fraction, field map quantification bias and susceptibility bias, and (c) evaluate the susceptibility mapping performance of the proposed effective in-phase echoes in comparison to single-peak in-phase echoes and water-fat separation for regions where both water and fat are present.

METHODS

Effective multipeak in-phase echo times for a bone marrow and a liver fat spectral model were derived from a single voxel simulation. A Monte Carlo simulation was performed to assess the field map estimation error as a function of fat fraction for the different in-phase echoes. Additionally, a phantom scan and in vivo scans in the liver, spine, and breast were performed and evaluated with respect to quantification accuracy.

RESULTS

The use of single-peak in-phase echoes can introduce a worst-case susceptibility bias of 0.43 $$ 0.43 $$  ppm. The use of effective multipeak in-phase echoes shows a similar quantitative performance in the numerical simulation, the phantom and in all in vivo anatomies when compared to water-fat separation-based QSM.

CONCLUSION

QSM based on the proposed effective multipeak in-phase echoes can alleviate the quantification bias present in QSM based on single-peak in-phase echoes. When compared to water-fat separation-based QSM the proposed effective in-phase echo times achieve a similar quantitative performance while drastically reducing the computational expense for field map estimation.

Globally optimal OCT surface segmentation using a constrained IPM optimization

Segmentation of multiple surfaces in optical coherence tomography (OCT) images is a challenging problem, further complicated by the frequent presence of weak boundaries, varying layer thicknesses, and mutual influence between adjacent surfaces. The traditional graph-based optimal surface segmentation method has proven its effectiveness with its ability to capture various surface priors in a uniform graph model. However, its efficacy heavily relies on handcrafted features that are used to define the surface cost for the "goodness" of a surface. Recently, deep learning (DL) is emerging as a powerful tool for medical image segmentation thanks to its superior feature learning capability. Unfortunately, due to the scarcity of training data in medical imaging, it is nontrivial for DL networks to implicitly learn the global structure of the target surfaces, including surface interactions. This study proposes to parameterize the surface cost functions in the graph model and leverage DL to learn those parameters. The multiple optimal surfaces are then simultaneously detected by minimizing the total surface cost while explicitly enforcing the mutual surface interaction constraints. The optimization problem is solved by the primal-dual interior-point method (IPM), which can be implemented by a layer of neural networks, enabling efficient end-to-end training of the whole network. Experiments on spectral-domain optical coherence tomography (SD-OCT) retinal layer segmentation demonstrated promising segmentation results with sub-pixel accuracy.

Segmentation techniques of brain arteriovenous malformations for 3D visualization: a systematic review

BACKGROUND

Visualization, analysis and characterization of the angioarchitecture of a brain arteriovenous malformation (bAVM) present crucial steps for understanding and management of these complex lesions. Three-dimensional (3D) segmentation and 3D visualization of bAVMs play hereby a significant role. We performed a systematic review regarding currently available 3D segmentation and visualization techniques for bAVMs.

METHODS

PubMed, Embase and Google Scholar were searched to identify studies reporting 3D segmentation techniques applied to bAVM characterization. Category of input scan, segmentation (automatic, semiautomatic, manual), time needed for segmentation and 3D visualization techniques were noted.

RESULTS

Thirty-three studies were included. Thirteen (39%) used MRI as baseline imaging modality, 9 used DSA (27%), and 7 used CT (21%). Segmentation through automatic algorithms was used in 20 (61%), semiautomatic segmentation in 6 (18%), and manual segmentation in 7 (21%) studies. Median automatic segmentation time was 10 min (IQR 33), semiautomatic 25 min (IQR 73). Manual segmentation time was reported in only one study, with the mean of 5-10 min. Thirty-two (97%) studies used screens to visualize the 3D segmentations outcomes and 1 (3%) study utilized a heads-up display (HUD). Integration with mixed reality was used in 4 studies (12%).

CONCLUSIONS

A golden standard for 3D visualization of bAVMs does not exist. This review describes a tendency over time to base segmentation on algorithms trained with machine learning. Unsupervised fuzzy-based algorithms thereby stand out as potential preferred strategy. Continued efforts will be necessary to improve algorithms, integrate complete hemodynamic assessment and find innovative tools for tridimensional visualization.

Spatio-temporal hybrid neural networks reduce erroneous human "judgement calls" in the diagnosis of Takotsubo syndrome

BACKGROUND

We investigate whether deep learning (DL) neural networks can reduce erroneous human "judgment calls" on bedside echocardiograms and help distinguish Takotsubo syndrome (TTS) from anterior wall ST segment elevation myocardial infarction (STEMI).

METHODS

We developed a single-channel (DCNN[2D SCI]), a multi-channel (DCNN[2D MCI]), and a 3-dimensional (DCNN[2D+t]) deep convolution neural network, and a recurrent neural network (RNN) based on 17,280 still-frame images and 540 videos from 2-dimensional echocardiograms in 10 years (1 January 2008 to 1 January 2018) retrospective cohort in University of Iowa (UI) and eight other medical centers. Echocardiograms from 450 UI patients were randomly divided into training and testing sets for internal training, testing, and model construction. Echocardiograms of 90 patients from the other medical centers were used for external validation to evaluate the model generalizability. A total of 49 board-certified human readers performed human-side classification on the same echocardiography dataset to compare the diagnostic performance and help data visualization.

FINDINGS

The DCNN (2D SCI), DCNN (2D MCI), DCNN(2D+t), and RNN models established based on UI dataset for TTS versus STEMI prediction showed mean diagnostic accuracy 73%, 75%, 80%, and 75% respectively, and mean diagnostic accuracy of 74%, 74%, 77%, and 73%, respectively, on the external validation. DCNN(2D+t) (area under the curve [AUC] 0·787 vs. 0·699, P = 0·015) and RNN models (AUC 0·774 vs. 0·699, P = 0·033) outperformed human readers in differentiating TTS and STEMI by reducing human erroneous judgement calls on TTS.

INTERPRETATION

Spatio-temporal hybrid DL neural networks reduce erroneous human "judgement calls" in distinguishing TTS from anterior wall STEMI based on bedside echocardiographic videos.

FUNDING

University of Iowa Obermann Center for Advanced Studies Interdisciplinary Research Grant, and Institute for Clinical and Translational Science Grant. National Institutes of Health Award (1R01EB025018-01).

Lipid droplet-size mapping in human adipose tissue using a clinical 3T system

PURPOSE

To develop a methodology for probing lipid droplet sizes with a clinical system based on a diffusion-weighted stimulated echo-prepared turbo spin-echo sequence and to validate the methodology in water-fat emulsions and show its applicability in ex vivo adipose-tissue samples.

METHODS

A diffusion-weighted stimulated echo-prepared preparation was combined with a single-shot turbo spin-echo readout for measurements at different b-values and diffusion times. The droplet size was estimated with an analytical expression, and three fitting approaches were compared: magnitude-based spatial averaging with voxel-wise residual minimization, complex-based spatial averaging with voxel-wise residual minimization, and complex-based spatial averaging with neighborhood-regularized residual minimization. Simulations were performed to characterize the fitting residual landscape and the approaches' noise performance. The applicability was assessed in oil-in-water emulsions in comparison with laser deflection and in ten human white adipose tissue samples in comparison with histology.

RESULTS

The fitting residual landscape showed a minimum valley with increasing extent as the droplet size increased. In phantoms, a very good agreement of the mean droplet size was observed between the diffusion-weighted MRI-based and the laser deflection measurements, showing the best performance with complex-based spatial averaging with neighborhood-regularized residual minimization processing (R² /P: 0.971/0.014). In the human adipose-tissue samples, complex-based spatial averaging with neighborhood-regularized residual minimization processing showed a significant correlation (R² /P: 0.531/0.017) compared with histology.

CONCLUSION

The proposed acquisition and parameter-estimation methodology was able to probe restricted diffusion effects in lipid droplets. The methodology was validated using phantoms, and its feasibility in measuring an apparent lipid droplet size was demonstrated ex vivo in white adipose tissue.

Improved body quantitative susceptibility mapping by using a variable-layer single-min-cut graph-cut for field-mapping

PURPOSE

To develop a robust algorithm for field-mapping in the presence of water-fat components, large B 0 field inhomogeneities and MR signal voids and to apply the developed method in body applications of quantitative susceptibility mapping (QSM).

METHODS

A framework solving the cost-function of the water-fat separation problem in a single-min-cut graph-cut based on the variable-layer graph construction concept was developed. The developed framework was applied to a numerical phantom enclosing an MR signal void, an air bubble experimental phantom, 14 large field of view (FOV) head/neck region in vivo scans and to 6 lumbar spine in vivo scans. Field-mapping and subsequent QSM results using the proposed algorithm were compared to results using an iterative graph-cut algorithm and a formerly proposed single-min-cut graph-cut.

RESULTS

The proposed method was shown to yield accurate field-map and susceptibility values in all simulation and in vivo datasets when compared to reference values (simulation) or literature values (in vivo). The proposed method showed improved field-map and susceptibility results compared to iterative graph-cut field-mapping especially in regions with low SNR, strong field-map variations and high R 2 ∗ values.

CONCLUSIONS

A single-min-cut graph-cut field-mapping method with a variable-layer construction was developed for field-mapping in body water-fat regions, improving quantitative susceptibility mapping particularly in areas close to MR signal voids.