Elastic Deformation of Optical Coherence Tomography Images of Diabetic Macular Edema for Deep-Learning Models Training: How Far to Go?

- Objective: To explore the clinical validity of elastic deformation of optical coherence tomography (OCT) images for data augmentation in the development of deep-learning model for detection of diabetic macular edema (DME). METHODS: Prospective evaluation of OCT images of DME (n = 320) subject to elastic transformation, with the deformation intensity represented by ([Formula: see text]). Three sets of images, each comprising 100 pairs of scans (100 original & 100 modified), were grouped according to the range of ([Formula: see text]), including low-, medium- and high-degree of augmentation; ([Formula: see text] = 1-6), ([Formula: see text] = 7-12), and ([Formula: see text] = 13-18), respectively. Three retina specialists evaluated all datasets in a blinded manner and designated each image as 'original' versus 'modified'. The rate of assignment of 'original' value to modified images (false-negative) was determined for each grader in each dataset. RESULTS: The false-negative rates ranged between 71-77% for the low-, 63-76% for the medium-, and 50-75% for the high-augmentation categories. The corresponding rates of correct identification of original images ranged between 75-85% ([Formula: see text]0.05) in the low-, 73-85% ([Formula: see text]0.05 for graders 1 & 2, p = 0.01 for grader 3) in the medium-, and 81-91% ([Formula: see text]) in the high-augmentation categories. In the subcategory ([Formula: see text] = 7-9) the false-negative rates were 93-83%, whereas the rates of correctly identifying original images ranged between 89-99% ([Formula: see text]0.05 for all graders). CONCLUSIONS: Deformation of low-medium intensity ([Formula: see text] = 1-9) may be applied without compromising OCT image representativeness in DME. Clinical and Translational Impact Statement-Elastic deformation may efficiently augment the size, robustness, and diversity of training datasets without altering their clinical value, enhancing the development of high-accuracy algorithms for automated interpretation of OCT images.

Automated 4-dimensional regional myocardial strain evaluation using cardiac computed tomography.

Evaluation of myocardial regional function is generally performed by visual "eyeballing" which is highly subjective. A robust quantifiable parameter of regional function is required to provide an objective, repeatable and comparable measure of myocardial performance. We aimed to evaluate the clinical utility of novel regional myocardial strain software from cardiac computed tomography (CT) datasets. 93 consecutive patients who had undergone retrospectively gated cardiac CT were evaluated by the software, which utilizes a finite element based tracking algorithm through the cardiac cycle. Circumferential (CS), longitudinal (LS) and radial (RS) strains were calculated for each of 16 myocardial segments and compared to a visual assessment, carried out by an experienced cardiologist on cine movies of standard "echo" views derived from the CT data. A subset of 37 cases was compared to speckle strain by echocardiography. The automated software performed successfully in 93/106 cases, with minimal human interaction. Peak CS, LS and RS all differentiated well between normal, hypokinetic and akinetic segments. Peak strains for akinetic segments were generally post-systolic, peaking at 50 ± 17% of the RR interval compared to 43 ± 9% for normokinetic segments. Using ROC analysis to test the ability to differentiate between normal and abnormal segments, the area under the curve was 0.84 ± 0.01 for CS, 0.80 ± 0.02 for RS and 0.68 ± 0.02 for LS. There was a moderate agreement with speckle strain. Automated 4D regional strain analysis of CT datasets shows a good correspondence to visual analysis and successfully differentiates between normal and abnormal segments, thus providing an objective quantifiable map of myocardial regional function.

Strain analysis from 4-D cardiac CT image data.

Strain is a discriminative parameter of regional myocardial dysfunction. Despite the large body of research on myocardial strain analysis in echocardiography and MR images, such techniques have not often been applied to cardiac CT data. Reasons for this include the challenges of sparse image deformation clues and the low temporal resolution. In the current study, we propose an algorithm that uses cardiac CT data to evaluate the mechanical function of the left ventricle. The algorithm is based on a deformable LV model that contains both the myocardium and the blood pool regions and that accounts for the elasticity and incompressibility of the myocardium with the rapid contraction of the blood pool. Our algorithm uses the image intensities of the trabeculle and papillary muscles as well as the border edges in an optical flow manner to extract the 3-D velocities. The resulting strains and rotational values derived from a set of normal patients correlate highly with values from the research literature. We validated our algorithm against 2-D speckle tracking analysis and against visual scores obtained by an expert. Our study shows that strain analysis using CT data can be used in clinical practice.

A neural network technique for remeshing of bone microstructure.

Today, there is major interest within the biomedical community in developing accurate noninvasive means for the evaluation of bone microstructure and bone quality. Recent improvements in 3D imaging technology, among them development of micro-CT and micro-MRI scanners, allow in-vivo 3D high-resolution scanning and reconstruction of large specimens or even whole bone models. Thus, the tendency today is to evaluate bone features using 3D assessment techniques rather than traditional 2D methods. For this purpose, high-quality meshing methods are required. However, the 3D meshes produced from current commercial systems usually are of low quality with respect to analysis and rapid prototyping. 3D model reconstruction of bone is difficult due to the complexity of bone microstructure. The small bone features lead to a great deal of neighborhood ambiguity near each vertex. The relatively new neural network method for mesh reconstruction has the potential to create or remesh 3D models accurately and quickly. A neural network (NN), which resembles an artificial intelligence (AI) algorithm, is a set of interconnected neurons, where each neuron is capable of making an autonomous arithmetic calculation. Moreover, each neuron is affected by its surrounding neurons through the structure of the network. This paper proposes an extension of the growing neural gas (GNN) neural network technique for remeshing a triangular manifold mesh that represents bone microstructure. This method has the advantage of reconstructing the surface of a genus-n freeform object without a priori knowledge regarding the original object, its topology, or its shape.