Ryan Amelon

Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset Through Integration of Deep Learning

Purpose To compare performance of a deep-learning enhanced algorithm for automated detection of diabetic retinopathy (DR), to the previously published performance of that algorithm, the Iowa Detection Program (IDP)-without deep learning components-on the same publicly available set of fundus images and previously reported consensus reference standard set, by three US Board certified retinal specialists. Methods We used the previously reported consensus reference standard of referable DR (rDR), defined as International Clinical Classification of Diabetic Retinopathy moderate, severe nonproliferative (NPDR), proliferative DR, and/or macular edema (ME). Neither Messidor-2 images, nor the three retinal specialists setting the Messidor-2 reference standard were used for training IDx-DR version X2.1. Sensitivity, specificity, negative predictive value, area under the curve (AUC), and their confidence intervals (CIs) were calculated. Results Sensitivity was 96.8% (95% CI: 93.3%-98.8%), specificity was 87.0% (95% CI: 84.2%-89.4%), with 6/874 false negatives, resulting in a negative predictive value of 99.0% (95% CI: 97.8%-99.6%). No cases of severe NPDR, PDR, or ME were missed. The AUC was 0.980 (95% CI: 0.968-0.992). Sensitivity was not statistically different from published IDP sensitivity, which had a CI of 94.4% to 99.3%, but specificity was significantly better than the published IDP specificity CI of 55.7% to 63.0%. Conclusions A deep-learning enhanced algorithm for the automated detection of DR, achieves significantly better performance than a previously reported, otherwise essentially identical, algorithm that does not employ deep learning. Deep learning enhanced algorithms have the potential to improve the efficiency of DR screening, and thereby to prevent visual loss and blindness from this devastating disease.

Three-dimensional characterization of regional lung deformation

The deformation of the lung during inspiration and expiration involves regional variations in volume change and orientational preferences. Studies have reported techniques for measuring the displacement field in the lung based on imaging or image registration. However, means of interpreting all the information in the displacement field in a physiologically relevant manner is lacking. We propose three indices of lung deformation that are determinable from the displacement field: the Jacobian--a measure of volume change, the anisotropic deformation index--a measure of the magnitude of directional preference in volume change and a slab-rod index--a measure of the nature of directional preference in volume change. To demonstrate the utility of these indices, they were determined for six human subjects using deformable image registration on static CT images, registered from FRC to TLC. Volume change was elevated in the inferior-dorsal region as should be expected for breathing in the supine position. The anisotropic deformation index was elevated in the inferior region owing to proximity to the diaphragm and in the lobar fissures owing to sliding. Vessel regions in the lung had a significantly rod-like deformation compared to the whole lung. Compared to upper lobes, lower lobes exhibited significantly greater volume change (19.4% and 21.3% greater in the right and left lungs on average; p<0.005) and anisotropy in deformation (26.3% and 21.8% greater in the right and left lungs on average; p<0.05) with remarkable consistency across subjects. The developed deformation indices lend themselves to exhaustive and physiologically intuitive interpretations of the displacement fields in the lung determined through image-registration techniques or finite element simulations.

A Measure for Characterizing Sliding on Lung Boundaries

The lobes of the lung slide relative to each other during breathing. Quantifying lobar sliding can aid in better understanding lung function, better modeling of lung dynamics, and for studying phenomenon such as pleural adhesion. We propose a novel measure to characterize lobe sliding in the lung based on the displacement field obtained from image registration of CT scans. When two sliding lobes are modeled as a continuum, the discontinuity in the displacement field at the fissure will manifest as elevated maximum shear—the proposed measure—which is capable of capturing both the level and orientation of sliding. Six human lungs were analyzed using scans spanning functional residual capacity to total lung capacity. The lung lobes were segmented and registered on a lobe-by-lobe basis to obtain the displacement field from which the proposed sliding measure was calculated. The sliding measure was found to be insignificant in the parenchyma, as relatively little tissue shear occurs here. On the other hand, it was elevated along the fissures. Thus, a map of the proposed sliding measure of the entire lung clearly delineates and quantifies sliding between lung lobes. Sliding is a key aspect of lung deformation during breathing. The proposed measure may help resolve artifacts introduced by sliding in deformation analysis techniques used for radiotherapy.

Unifying vascular information in intensity-based nonrigid lung CT registration

Image registration plays an important role within pulmonary image analysis. Accurate registration is critical to post-analysis of lung mechanical properties. To improve registration accuracy, we utilize the rich information of vessel locations and shapes, and introduce a new similarity criterion, sum of squared vesselness measure difference (SSVMD). This metric is added to three existing intensity-based similarity criteria for nonrigid lung CT image registration to show its ability in improving matching accuracy. The registration accuracy is assessed by landmark error calculation and distance map visualization on vascular tree. The average landmark errors are reduced by over 20% and are within 0.7 mm after adding SSVMD constraint to three existing intensity-based similarity metrics. Visual inspection shows matching accuracy improvements in the lung regions near the thoracic cage and near the diaphragm. Experiments also show this vesselness constraint makes the Jacobian map of transformations physiologically more plausible and reliable.

High bone density masks architectural deficiencies in an individual with spinal cord injury.

Abstract Context Spinal cord injury (SCI) causes a decline of bone mineral density (BMD) in the paralyzed extremities via the gradual degradation and resorption of trabecular elements. Clinical tools that report BMD may not offer insight into trabecular architecture flaws that could affect bones ability to withstand loading. We present a case of a woman with a 30-year history of SCI and abnormally high distal femur BMD. Findings Peripheral quantitative-computed tomography-based BMD for this subject was ∼20% higher than previously published non-SCI values. Computed tomography (CT) revealed evidence of sclerotic bone deposition in the trabecular envelope, most likely due to glucocorticoid-induced osteonecrosis. Volumetric topologic analysis of trabecular architecture indicated that the majority of the bone mineral was organized into thick, plate-like structures rather than a multi-branched trabecular network. Visual analysis of the CT stack confirmed that the sclerotic bone regions were continuous with the cortex at only a handful of points. Conclusions Conventional clinical BMD analysis could have led to erroneous assumptions about this subjects bone quality. CT-based analysis revealed that this subjects high BMD masked underlying architectural flaws. For patients who received prolonged glucocorticoid therapy, excessively high BMD should be viewed with caution. The ability of this subjects bone to resist fracture is, in our view, extremely suspect. A better understanding of the mechanical competency of this very dense, but architecturally flawed bone would be desirable before this subject engaged in activities that load the limbs.

Assessment of trabecular bone strength at in vivo CT imaging with space-variant hysteresis and finite element modelling

Osteoporosis is a common bone disease associated with reduced bone strength and increased fracture risk. Finite element modelling (FEM) is a powerful tool to assess bone strength. Fast acquisition and highly reduced radiation in latest multi-row detector CT (MDCT) put it as a frontline imaging modality to assess in vivo trabecular bone (TB) microarchitecture and strength. In current CT image resolution, conventional segmentation methods fail to maintain TB network connectivity limiting reliable assessment of TB strength using FEM. In this paper, we present a new space-variant hysteresis approach to maintain TB connectivity while preserving marrow pores and a high quality mesh generator for FEM. We examine the effectiveness in estimating TB strength. The reproducibility and the ability to predict actual bone strength were examined on MDCT images of cadaveric ankle specimens under in vivo conditions. An intra-class correlation coefficient of 0.97 was observed in computed Youngs modulus from repeat scans, and a high linear correlation (R2=0.92) were found between computed and experimental Youngs modulus.

Intensity-Based Registration for Lung Motion Estimation

Image registration plays an important role within pulmonary image analysis. The task of registration is to find the spatial mapping that brings two images into alignment. Registration algorithms designed for matching 4D lung scans or two 3D scans acquired at different inflation levels can catch the temporal changes in position and shape of the region of interest. Accurate registration is critical to post-analysis of lung mechanics and motion estimation. In this chapter, we discuss lung-specific adaptations of intensity-based registration methods for 3D/4D lung images and review approaches for assessing registration accuracy. Then we introduce methods for estimating tissue motion and studying lung mechanics. Finally, we discuss methods for assessing and quantifying specific volume change, specific ventilation, strain/ stretch information and lobar sliding.

Tracking regional tissue volume and function change in lung using image registration

We have previously demonstrated the 24-hour redistribution and reabsorption of bronchoalveolar lavage (BAL) fluid delivered to the lung during a bronchoscopic procedure in normal volunteers. In this work we utilize image-matching procedures to correlate fluid redistribution and reabsorption to changes in regional lung function. Lung CT datasets from six human subjects were used in this study. Each subject was scanned at four time points before and after BAL procedure. Image registration was performed to align images at different time points and different inflation levels. The resulting dense displacement fields were utilized to track tissue volume changes and reveal deformation patterns of local parenchymal tissue quantitatively. The registration accuracy was assessed by measuring landmark matching errors, which were on the order of 1 mm. The results show that quantitative-assessed fluid volume agreed well with bronchoscopist-reported unretrieved BAL volume in the whole lungs (squared linear correlation coefficient was 0.81). The average difference of lung tissue volume at baseline and after 24 hours was around 2%, which indicates that BAL fluid in the lungs was almost absorbed after 24 hours. Regional lung-function changes correlated with the presence of BAL fluid, and regional function returned to baseline as the fluid was reabsorbed.

Estimation of Lung Ventilation

Since the primary function of the lung is gas exchange, ventilation can be interpreted as an index of lung function in addition to perfusion. Injury and disease processes can alter lung function on a global and/or a local level. MDCT can be used to acquire multiple static breath-hold CT images of the lung taken at different lung volumes, or with proper respiratory control, 4DCT images of the lung reconstructed at different respiratory phases. Image registration can be applied to this data to estimate a deformation field that transforms the lung from one volume configuration to the other. This deformation field can be analyzed to estimate local lung tissue expansion, calculate voxel-by-voxel intensity change, and make biomechanical measurements. The physiologic significance of the registration-based measures of respiratory function can be established by comparing to more conventional measurements, such as nuclear medicine or contrast wash-in/wash-out studies with CT or MR. An important emerging application of these methods is the detection of pulmonary function change in subjects undergoing radiation therapy (RT) for lung cancer. During RT, treatment is commonly limited to sub-therapeutic doses due to unintended toxicity to normal lung tissue. Measurement of pulmonary function may be useful as a planning tool during RT planning, may be useful for tracking the progression of toxicity to nearby normal tissue during RT, and can be used to evaluate the effectiveness of a treatment post-therapy. This chapter reviews the basic measures to estimate regional ventilation from image registration of CT images, the comparison of them to the existing golden standard and the application in radiation therapy.

A Novel Method of Characterizing Regional Lung Deformation

Regional lung deformation (as opposed to whole lung volume change) may be an indicator of localized lung malformations. Regional deformation may be characterized by principal strains, but lack a direct physiological relevance. Alternatively, regional volume change (distribution of Jacobian) is physiologically intuitive, but does not characterize all aspects of deformation. For instance, a region may undergo no volume change, but still have deformed significantly — say, when the lengthening in one direction is compensated by contraction along another direction. From a perspective of physiological relevance to lung function, lung deformation may be thought to encompass both the volume change and directional preferences in volume change.Copyright