Region Growing Image Segmentation Algorithms Research Articles

Identification of Nodular Lymphocyte-Predominant Hodgkin Lymphoma Variant Morphology Using Artificial Intelligence

Background Lymphocyte predominant (LP) cells located outside B-cell-rich nodules are the hallmark of variant morphologic patterns of nodular lymphocyte predominant Hodgkin lymphoma (NLPHL). In comparison with typical NLPHL (patterns A and B), cases with variant morpholoy (patterns C, D, E and F) are associated with significantly worse clinical behavior. Nevertheless, assessment of tissue architecture may be challenging in limited biopsy specimens, which are increasingly common in routine clinical practice. We sought to address this gap by developing artificial intelligence-assisted modules to accurately recognize NLPHL and distinguish patterns A/B from variant patterns, namely the most common of those, pattern C. Material & Methods The study group included whole slide images (WSI) of hematoxylin and eosin (H&E) stains and CD20 immunohistochemistry stains from 10 NLPHL patients; cases emcompassed patterns A, B & C (A/B: 5 cases; C: 5 cases) as determined by hematopathology review. We designed modules to assess the spatial relationship between LP cells and B cell-rich areas by first seeking to detect LP cells using high-power field (HPF) magnification of H&E WSI (Figure A1). We split the WSI into tiles (8000 x 8000 pixels) to mitigate the computational burden of WSI analysis. For each tile, we deployed the pretrained HoVer-Net for nuclear segmentation. To distinguish LP cells from surrounding small lymphocytes, we analyzed nuclear size, intensity, and gray-level co-courrance matrix (GLCM) features (correlation and dissimilarity). For each feature, we identified the optimal threshold based on feature histogram of LP in comparison to the remainder surrounding cells. Nuclei meeting all feature criteria were labeled as LP cells and validated morphologically by hematopathologists. All tiles were then stitched to illustrate the spatial distribution of LP cells on WSI. Next we deployed the active contour (AC) technique, an iterative region-growing image segmentation algorithm, to identify B-cell areas from CD20 stains (Figure A2). To reliably define the intial contours for the AC, we abided by the following criteria: 1) Red (R) channel intensity larger than both the green (G) and blue (B) channels, with a ratio of 1.25; 2) Including pixels in the bottom 10% of the intensity histogram after removal of slide background and converting the RGB image into gray scale. We merged the contours from these two techniques and used them as initial seeds for AC segmentation, after which we took the final evolved contour as the boundaries of B-cell areas. We then overlayed CD20 on H&E WSI to assess the spatial relationship between the detected LP cells and B-cell areas. Since tissue sections on the CD20 and H&E slides represent neighboring slices, only translation, rotation, and slight scaling were surmised to occurr between slides in this pilot study set. Accordingly, we employed the affine registration paradigm and optimized mutual information between the transformed CD20 and fixed H&E images until convergence then warped the detected B-cell areas via the obtained deformation field (Figure A3). Results The concordant positioning between two stains after registration allowed us to compare the spatial relationship of LP cells to B-cell areas (including assessment of the number of LP cells inside (922) and outside (639) B-cell areas) (Figure A4). We then segmented the final image into quadrants to simulate needle core biopsies and confirmed the efficiency of our modules in detecting LP cells and analyzing their relationship to B-cell areas in limited tissue, in all quadrants (Figure A5). Our model was successfully applied on 100% of 10 NLPHL cases in this pilot group. After proof of concept was established in a full tissue biopsy sample (Figure A1-5), we applied the same steps to an actual physical needle core biopsy of an NLPHL-pattern C case (Figure B1-3) and were able to identify 7 out of 29 LP cells located outside B-cell areas, in congruence with the previously rendered diagnosis (Figure B4). Conclusion This technique, using H&E staining and CD20 immunohistochemistry only, is especially promising to work-up cases when ordering a routine panel of immunohistochemical stains is not feasible due to limited tissue that may be exhausted. Further validation using an independent cohort is warranted in the future. Figure 1View largeDownload PPTFigure 1View largeDownload PPT Close modal

Adaptive Region Growing Image Segmentation Algorithms for Breast MRI

Early detection and characterization of breast lesion are important for a better and effective treatment of breast cancer. In this paper, four different adaptive region growing image segmentation algorithms are compared. In fact, seed selection was a vital step in the success of region growing methods, so, better schemes for seed selection methods are proposed, namely, joint probabilistic seed selection (JPSS) and Generalised simulated annealing (GSA) based seed selection. The proposed region growing methods namely Fuzzy Region Growing (FRG) and Neutrosophic Region Growing (NRG) are integrated as JPSS-FRG and GSA-NRG frameworks. Another two methods are Scale Invariant Region growing (SiRG) and Fuzzy Neutrosophic Confidence Region growing (FNCRG). The results showed that FNCRG algorithm increases breast cancer detection rate on MRI breast images with the maximum of 93% is achieved. SiRG algorithm improves the true positive rate by 13% compared to existing methods. Further, GSA-NRG makes better segmentation accuracy by 9% and true positive rate by 12%. Also, JPSS-FRG algorithm enhances segmentation accuracy by 24% and improving the true positive rate by 27% compared to Region Growing-Cellular Neural Network (RG-CNN) and Seeded Region Growing-Particle swarm optimization (SRG-PSO) methods respectively.

SYNCHRONIC CORONAL HOLE MAPPING USING MULTI-INSTRUMENT EUV IMAGES: DATA PREPARATION AND DETECTION METHOD

ABSTRACT We describe a method for the automatic mapping of coronal holes (CHs) using simultaneous multi-instrument EUV imaging data. Synchronized EUV images from STEREO/EUVI A and B 195 Å and SDO/AIA 193 Å are preprocessed, including point-spread function deconvolution and the application of data-derived intensity corrections which account for center-to-limb variations (limb brightening) and inter-instrument intensity normalization. We systematically derive a robust limb-brightening correction that takes advantage of unbiased long-term averages of data and respects the physical nature of the problem. The new preprocessing greatly assists in CH detection, allowing for the use of a simplified variable-connectivity, two-threshold region-growing image segmentation algorithm to obtain consistent detection results. We generate synchronic EUV and CH maps, and show a preliminary analysis of CH evolution. Several data and code products are made available to the community (www.predsci.com/chd): for the period of this study (2010 June 10 to 2014 August 18), we provide synchronic EUV and CH map data at 6 hr cadence, data-derived limb-brightening corrections for STEREO/EUVI A and B 195 Å and SDO/AIA 193 Å, and inter-instrument correction factors to equate their intensities. We also provide the CH image segmentation code module (ezseg) implemented in both FORTRAN-OpenMP and GPU-accelerated C-CUDA. A complete implementation of our CH detection pipeline in the form of a ready-to-use MATLAB driver script euv2chm utilizing ezseg is also made available.

Automated mapping of burned areas in semi-arid ecosystems using modis time-series imagery

Abstract. Understanding spatial and temporal patterns of burned areas at regional scales, provides a long-term perspective of fire processes and its effects on ecosystems and vegetation recovery patterns, and it is a key factor to design prevention and post-fire restoration plans and strategies. Standard satellite burned area and active fire products derived from the 500-m MODIS and SPOT are avail - able to this end. However, prior research caution on the use of these global-scale products for regional and sub-regional applica - tions. Consequently, we propose a novel algorithm for automated identification and mapping of burned areas at regional scale in semi-arid shrublands. The algorithm uses a set of the Normalized Burned Ratio Index products derived from MODIS time series; using a two-phased cycle, it firstly detects potentially burned pixels while keeping a low commission error (false detection of burned areas), and subsequently labels them as seed patches. Region growing image segmentation algorithms are applied to the seed patches in the second-phase, to define the perimeter of fire affected areas while decreasing omission errors (missing real burned areas). Independently-derived Landsat ETM+ burned-area reference data was used for validation purposes. The correlation between the size of burnt areas detected by the global fire products and independently-derived Landsat reference data ranged from R2 = 0.01 - 0.28, while our algorithm performed showed a stronger correlation coefficient (R2 = 0.96). Our findings confirm prior research calling for caution when using the global fire products locally or regionally.

Semi-automated mapping of burned areas in semi-arid ecosystems using MODIS time-series imagery

Understanding spatial and temporal patterns of burned areas at regional scales, provides a long-term perspective of fire processes and its effects on ecosystems and vegetation recovery patterns, and it is a key factor to design prevention and post-fire restoration plans and strategies. Remote sensing has become the most widely used tool to detect fire affected areas over large tracts of land (e.g., ecosystem, regional and global levels). Standard satellite burned area and active fire products derived from the 500-m Moderate Resolution Imaging Spectroradiometer (MODIS) and the Satellite Pour l’Observation de la Terre (SPOT) are available to this end. However, prior research caution on the use of these global-scale products for regional and sub-regional applications. Consequently, we propose a novel semi-automated algorithm for identification and mapping of burned areas at regional scale. The semi-arid Monte shrublands, a biome covering 240,000km2 in the western part of Argentina, and exposed to seasonal bushfires was selected as the test area. The algorithm uses a set of the normalized burned ratio index products derived from MODIS time series; using a two-phased cycle, it firstly detects potentially burned pixels while keeping a low commission error (false detection of burned areas), and subsequently labels them as seed patches. Region growing image segmentation algorithms are applied to the seed patches in the second-phase, to define the perimeter of fire affected areas while decreasing omission errors (missing real burned areas). Independently-derived Landsat ETM+ burned-area reference data was used for validation purposes. Additionally, the performance of the adaptive algorithm was assessed against standard global fire products derived from MODIS Aqua and Terra satellites, total burned area (MCD45A1), the active fire algorithm (MOD14); and the L3JRC SPOT VEGETATION 1km GLOBCARBON products. The correlation between the size of burned areas detected by the global fire products and independently-derived Landsat reference data ranged from R2=0.01–0.28, while our algorithm performed showed a stronger correlation coefficient (R2=0.96). Our findings confirm prior research calling for caution when using the global fire products locally or regionally.

A Region-Growing Segmentation Algorithm for GPUs

This letter proposes a parallel version for graphics processing units (GPU) of a region-growing image segmentation algorithm widely used by the geographic object-based image analysis (GEOBIA) community. Initially, all image pixels are considered as seeds or primitive segments. Fine-grained parallel threads assigned to individual pixels merge adjacent segments iteratively always ensuring to minimize the overall heterogeneity increase. Besides spectral features the merging criterion considers morphological features that can be efficiently computed in the underlying GPU architecture. Two alternatives using different merging criteria are proposed and tested. An experimental analysis upon five different test images has shown that the parallel algorithm may run up to 19 times faster than its sequential counterpart.

Optimal region growing segmentation and its effect on classification accuracy

Image segmentation is a preliminary and critical step in object-based image classification. Its proper evaluation ensures that the best segmentation is used in image classification. In this article, image segmentations with nine different parameter settings were carried out with a multi-spectral Landsat imagery and the segmentation results were evaluated with an objective function that aims at maximizing homogeneity within segments and separability between neighbouring segments. The segmented images were classified into eight land-cover classes and the classifications were evaluated with independent ground data comprising 600 randomly distributed points. The accuracy assessment results presented similar distribution as that of the objective function values, that is segmentations with the highest objective function values also resulted in the highest classification accuracies. This result shows that image segmentation has a direct effect on the classification accuracy; the objective function not only worked on a single band image as proved by (Espindola, G.M., Camara, G., Reis, I.A., Bins, L.S. and Monteiro, A.M., 2006, Parameter selection for region-growing image segmentation algorithms using spatial autocorrelation. International Journal of Remote Sensing, 27, pp. 3035–3040.) but also on multi-spectral imagery as tested in this, and is indeed an effective way to determine the optimal segmentation parameters. McNemar's test (z 2 = 10.27) shows that with the optimal segmentation, object-based classification achieved accuracy significantly higher than that of the pixel-based classification, with 99% significance level.

The load unbalancing problem for region growing image segmentation algorithms

This paper discusses and evaluates parallel implementations of a segmentation algorithm based on the Split-and-Merge approach. The solution has been conceived for a multiprocessor using the Single Program Multiple Data (SPMD) programming model and executions have been carried out on a Cray-T3E system. Our main goal is to describe our experiences in solving the region growing problem, which is representative of a class of non-uniform problems, characterized by a behavior that is data dependent. Since this problem exhibits unpredictable load fluctuations, it requires the use of load-balancing schemes to achieve efficient parallel solutions. We also propose and analyze several strategies for the selection of region identifiers and their influence on execution time and load distribution.