科技英语阅读与写作

Noha El-Zehiry1, Oliver Hayden2 and Ali Kamen1
1
Medical Imaging Technologies, Siemens Healthcare, Princeton, NJ, USA 2 In-Vitro DX and Bioscience, Siemens Healthcare, Erlangen, Germany
Fig. 1. Pipeline of Cell Extraction Algorithm. Left: original image superimposed by cell center. Middle: Result of aggregation of the probability map response and generating the internal markers. Left: Segmentation results.
tools or utilizegeneric cell segmentation tools [4] (that) are not designed forDHM and the accuracy of their performance for DHM wasnot assessed in the literature. One notable exception is the cellsegmentation approach presented by Yi et al.in [3]. The algorithm in [3] uses a sequence of morphological operations onthe phase image to generate markers for maker-controlled watershed segmentation. The approach is relatively complicatedand the main disadvantage is its sensitivity to the parametersof the morphological operators such as the size of the structure element used in every operation. We will overcome thisproblem. In this paper, we present a two-step segmentationapproach: Cell detection to localize the centers of the cellsand cell segmentation to delineate the boundaries of the cells.Qualitative and quantitative assessment of our segmentationmethod is presented. Moreover, we introduce a comparison tothe state-of-the art cell segmentation methods [3] and [4]. (4)The rest of the paper is organizedas follows: Section 2presents the details of our proposed segmentation methods,section 3 introduces the experimental results and the comparison to other segmentation methods, then a concludingdiscussion is provided in section4.
ABSTRACT Digital Holographic Microscopy (DHM) is becoming recently very popular for cell imaging. ① (1) The main advantageof digital holographic microscopy over classical microscopytechniques is that it does not only provide the projected imageof the object but also provides three dimensional informationof the object’s optical thickness. DHM technology could bethe core of a label-free imaging for hematology applications.In an ideal framework, a blood sample can be imaged usingDHM, machine learning approaches can (be used for) the cellextraction, differentiation and consequently computing all therelevant blood statistics such as the Mean Corpuscular Volume (MCV), the Red Blood Cell (RBC) count, Red BloodCell Distribution Width (RDW).The most vital componentin such a framework is accurate extraction of the cells. ② (2)Thispaper presents a novel approach to cell segmentation in whicha probabilistic boosting tree classifier is trained to detect thecenters of the cells using Haar-Features. The detected cellcenters are used to trigger a marker-controlled power watershed segmentation to compute the cell boundaries. Additionally, we present a comprehensive evaluation of segmentationmethods for cell extraction in digital holographic images. 1. INTRODUCTION Digital Holographic Microscopy (DHM) has received a lot ofattention recently in cell imaging [1, 2, 3].(3)DHM, if associated with proper image processing algorithms, can serve as the fundamental building block in a label free cell diagnostics workflow. In such a workflow, accuratecell extraction is a vital step to perform the analysis. Therefore, thorough investigation of cell segmentation in DHMimages is a persistent necessity. Current DHM applicationstudies such as [1, 2] use simple segmentation
that could work efficiently for all the images. Onthe other hand, tuning the parameters for each single imageis completely impractical and negates the high throughput advantage associated with DHM technology. Motivated by theprevious drawbacks, we present a new cell segmentation approach. The novelty of our approach is two-fold: First, robustmarker generation (using) cell center detection. ④ (6)Second, weuse power watershed for the cell segmentation whichhas beenprovenmore efficient than conventional watershed in generalsegmentation problems [7]. Figure 1 shows the pipeline of ourcell extraction approach. 2.1.Robust Marker Generation (7)We consider the marker generation problem as an object detection problem where we aim at finding the positions of cellcenters. For this purpose, we use a machine learning basedapproach instead of morphological operations to minimize thesensitivity to parameters’ choice. In this context, we use Probabilistic Boosting Tree (PBT) learning framework [8]. In thelearning phase, the PBT constructs a tree (in which) each nodecombines a set of weak classifiers into a strong classifier. Intesting phase, the conditional probability is computed at eachnode and the probability is propagated to the source of thetree to provide the overall probability. In our training, we usethe Haar features [9] to form the weak classifiers. ⑤ (8)In testing,we compute the probability ofeach pixel in the image beinga center of a cell, the probability map is thresholded to keeponly the pixels that are more likely to be a cell center. 2.2. Aggregation of the Detection Responses After computing the probability map and applying the threshold, we get a high response inside each cell. However, theresponse is not necessarily smooth and connected which maylead to false identification of one cell as multiple cells. Therefore, to aggregate these response, we apply a clustering stepon the thresholded probability map. The clustering serves twopurposes, first, it aggregated the cell responses in a single cell.Second, it provides a larger set of pixels to form as the internal marker for the segmentation. The pixels of each clusterare merged into a