Machine Learning-Based Segmentation of Left Ventricular Myocardial Fibrosis from Magnetic Resonance Imaging.

PURPOSE OF REVIEW: Myocardial fibrosis (MF) arises due to myocardial infarction and numerous cardiac diseases. MF may lead to several heart disorders, such as heart failure, arrhythmias, and ischemia. Cardiac magnetic resonance (CMR) imaging techniques, such as late gadolinium enhancement (LGE) CMR, enable non-invasive assessment of MF in the left ventricle (LV). Manual assessment of MF on CMR is a tedious and time-consuming task that is subject to high observer variability. Automated segmentation and quantification of MF is important for risk stratification and treatment planning in patients with heart disorders. This article aims to review the machine learning (ML)-based methodologies developed for MF quantification in the LV using CMR images. RECENT FINDINGS: With the availability of relatively large labeled datasets supervised learning methods based on both conventional ML and state-of-the-art deep learning (DL) methods have been successfully applied for automated segmentation of MF. The incorporation of ML algorithms into imaging techniques such as 3D LGE CMR permits fast characterization of MF on CMR imaging and may enhance the diagnosis and prognosis of patients with heart disorders. Concurrently, the studies using cine CMR images have revealed that accurate segmentation of MF on non-contrast CMR imaging might be possible. The application of ML/DL tools in CMR image interpretation is likely to result in accurate and efficient quantification of MF.

Automated segmentation of villi in histopathology images of placenta.

PURPOSE: Manual analysis of clinical placenta pathology samples under the microscope is a costly and time-consuming task. Computer-aided diagnosis might offer a means to obtain fast and reliable results and also substantially reduce inter- and intra-rater variability. Here, we present a fully automated segmentation method that is capable of distinguishing the complex histological features of the human placenta (i.e., the chorionic villous structures). METHODS: The proposed pipeline consists of multiple steps to segment individual placental villi structures in hematoxylin and eosin (H&E) stained placental images. Artifacts and undesired objects in the histological field of view are detected and excluded from further analysis. One of the challenges in our new algorithm is the detection and segmentation of touching villi in our dataset. The proposed algorithm uses the top-hat transformation to detect candidate concavities in each structure, which might represent two distinct villous structures in close proximity. The detected concavities are classified by extracting multiple features from each candidate concavity. Our proposed pipeline is evaluated against manual segmentations, confirmed by an expert pathologist, on 12 scans from three healthy control patients and nine patients diagnosed with preeclampsia, containing nearly 5000 individual villi. The results of our method are compared to a previously published method for villi segmentation. RESULTS: Our algorithm detected placental villous structures with an F1 score of 80.76% and sensitivity of 82.18%. These values are substantially better than the previously published method, whose F1 score and sensitivity are 65.30% and 55.12%, respectively. CONCLUSION: Our method is capable of distinguishing the complex histological features of the human placenta (i.e., the chorionic villous structures), removing artifacts over a large histopathology sample of human placenta, and (importantly) account for touching adjacent villi structures. Compared to existing methods, our developed method yielded high accuracy in detecting villi in placental images.

Three-dimensional segmentation of three-dimensional ultrasound carotid atherosclerosis using sparse field level sets.

PURPOSE: Three-dimensional ultrasound (3DUS) vessel wall volume (VWV) provides a 3D measurement of carotid artery wall remodeling and atherosclerotic plaque and is sensitive to temporal changes of carotid plaque burden. Unfortunately, although 3DUS VWV provides many advantages compared to measurements of arterial wall thickening or plaque alone, it is still not widely used in research or clinical practice because of the inordinate amount of time required to train observers and to generate 3DUS VWV measurements. In this regard, semiautomated methods for segmentation of the carotid media-adventitia boundary (MAB) and the lumen-intima boundary (LIB) would greatly improve the time to train observers and for them to generate 3DUS VWV measurements with high reproducibility. METHODS: The authors describe a 3D algorithm based on a modified sparse field level set method for segmenting the MAB and LIB of the common carotid artery (CCA) from 3DUS images. To the authors' knowledge, the proposed algorithm is the first direct 3D segmentation method, which has been validated for segmenting both the carotid MAB and the LIB from 3DUS images for the purpose of computing VWV. Initialization of the algorithm requires the observer to choose anchor points on each boundary on a set of transverse slices with a user-specified interslice distance (ISD), in which larger ISD requires fewer user interactions than smaller ISD. To address the challenges of the MAB and LIB segmentations from 3DUS images, the authors integrated regional- and boundary-based image statistics, expert initializations, and anatomically motivated boundary separation into the segmentation. The MAB is segmented by incorporating local region-based image information, image gradients, and the anchor points provided by the observer. Moreover, a local smoothness term is utilized to maintain the smooth surface of the MAB. The LIB is segmented by constraining its evolution using the already segmented surface of the MAB, in addition to the global region-based information and the anchor points. The algorithm-generated surfaces were sliced and evaluated with respect to manual segmentations on a slice-by-slice basis using 21 3DUS images. RESULTS: The authors used ISD of 1, 2, 3, 4, and 10 mm for algorithm initialization to generate segmentation results. The algorithm-generated accuracy and intraobserver variability results are comparable to the previous methods, but with fewer user interactions. For example, for the ISD of 3 mm, the algorithm yielded an average Dice coefficient of 94.4% ± 2.2% and 90.6% ± 5.0% for the MAB and LIB and the coefficient of variation of 6.8% for computing the VWV of the CCA, while requiring only 1.72 min (vs 8.3 min for manual segmentation) for a 3DUS image. CONCLUSIONS: The proposed 3D semiautomated segmentation algorithm yielded high-accuracy and high-repeatability, while reducing the expert interaction required for initializing the algorithm than the previous 2D methods.

Three-dimensional ultrasound of carotid atherosclerosis: semiautomated segmentation using a level set-based method.

PURPOSE: Three-dimensional ultrasound (3D US) of the carotid artery provides measurements of arterial wall and plaque [vessel wall volume (VWV)] that are complementary to the one-dimensional measurement of the carotid artery intima-media thickness. 3D US VWV requires an observer to delineate the media-adventitia boundary (MAB) and lumen-intima boundary (LIB) of the carotid artery. The main purpose of this work was to develop and evaluate a semiautomated segmentation algorithm for delineating the MAB and LIB of the carotid artery from 3D US images. METHODS: To segment the MAB and LIB, the authors used a level set method and combined several low-level image cues with high-level domain knowledge and limited user interaction. First, the operator initialized the algorithm by choosing anchor points on the boundaries, identified in the images. The MAB was segmented using local region- and edge-based energies and an energy that encourages the boundary to pass through anchor points from the preprocessed images. For the LIB segmentation, the authors used local and global region-based energies, the anchor point-based energy, as well as a constraint promoting a boundary separation between the MAB and LIB. The data set consisted of 231 2D images (11 2D images per each of 21 subjects) extracted from 3D US images. The image slices were segmented five times each by a single observer using the algorithm and the manual method. Volume-based, region-based, and boundary distance-based metrics were used to evaluate accuracy. Moreover, repeated measures analysis was used to evaluate precision. RESULTS: The algorithm yielded an absolute VWV difference of 5.0% +/- 4.3% with a segmentation bias of -0.9% +/- 6.6%. For the MAB and LIB segmentations, the method gave absolute volume differences of 2.5% +/- 1.8% and 5.6% +/- 3.0%, Dice coefficients of 95.4% +/- 1.6% and 93.1% +/- 3.1%, mean absolute distances of 0.2 +/- 0.1 and 0.2 +/- 0.1 mm, and maximum absolute distances of 0.6 +/- 0.3 and 0.7 +/- 0.6 mm, respectively. The coefficients of variation of the algorithm (5.1%) and manual methods (3.9%) were not significantly different, but the average time saved using the algorithm (2.8 min versus 8.3 min) was substantial. CONCLUSIONS: The authors generated and tested a semiautomated carotid artery VWV measurement tool to provide measurements with reduced operator time and interaction, with high Dice coefficients, and with necessary required precision.