To the Editor.—We read with great interest the protocols for the examination of specimens from patients with neuroblastoma and related neuroblastic tumors by Qualman et al.1 The authors suggested that pathologists follow the presented protocols when writing pathology reports on neuroblastoma and related neuroblastic tumors. The article introduces various essential elements of protocols, such as background documentation, explanatory notes, differential diagnosis, and a neuroblastoma staging system. These protocols are designed to assist pathologists in providing clinically useful and relevant information when reporting results of specimen examinations. To further assist pathologists, we are developing computerized techniques that can analyze whole slide tissue images and produce clinically useful information for neuroblastoma prognosis based on the International Neuroblastoma Classification System developed by Shimada et al.2The current prognostic evaluation for neuroblastoma patients requires microscopic examination of tumor tissue specimens for identifying certain morphologic characteristics. Because this visual evaluation process is inevitably subjective, the prognosis decision often suffers from interreviewer and intrareviewer variability. A recent study reports that there is a 20% discrepancy between evaluations from central and institutional reviewers.3With the advance in technology of digital scanners, it is now feasible to scan neuroblastoma tissue specimens and acquire whole slide digital images. This, in turn, allows us to devise a computer-aided classification methodology that can generate key quantifiable parameters useful for prognostic evaluation. A computerized system can serve as a second reader that assists pathologists in their evaluations to help improve the objectivity of the prognosis. For this purpose, we have developed and tested a computerized system to analyze neuroblastoma slides and report neuroblastic differentiation in percentages of undifferentiated, poorly differentiated, and differentiating regions of the slide under consideration.With our image analysis system, each tumor image is first segmented into multiple cytologic components using a novel, automated segmentation method that we developed. Discriminating features (ie, quantized measures of the physical, histopathologic, and statistical characteristics) are extracted from the segmented image regions. Next, feature values are inputted into a family of statistical classifiers (ie, parametric functions) trained with sample images as well as clinical knowledge about neuroblastoma. In our implementation, all image features are derived from color information and texture. Texture is characterized as the spatial variation of image intensities. The clinical information includes the characteristics of red blood cells, relevant principal cytologic components such as nuclei, cytoplasm, and neuropil, and size and shape of these components in relationship to each other.To achieve good computational efficiency, our system uses a multiple-resolution framework to emulate the way pathologists examine histology slides. The designed algorithm initially examines the images at the lowest resolution (similar to a low optical magnification view with a microscope). Because the lowest-resolution images are the smallest, it takes the least amount of time to process these images using the image analysis software. However, if the images at lower-resolution levels do not contain sufficient image details, the computerized system automatically switches to work on images of the next higher resolution (similar to a higher optical magnification view with a microscope).In our study, all neuroblastoma tumor slides were collected from Columbus Nationwide Children's Hospital (Columbus, Ohio) in compliance with an institutional review board protocol. According to the protocols commonly used in the Children's Oncology Group, these tissue slides were cut at a thickness of 5 μm and stained with hematoxylin-eosin. Each tissue slide was then digitized using a ScanScope T2 digitizer (Aperio, San Diego, Calif) at ×40 magnification. The resulting digital images were compressed following the JPEG compression standards at a compression ratio of approximately 1:40. After the compression, the typical image sizes can vary from 1 to about 4 GB with a typical spatial resolution of 60 000 × 60 000 in pixels.The image data set used in this study consists of 36 neuroblastoma cases, covering all 3 subtypes of neuroblastic grading. In our study, the training data set consisted of images of 3 representative cases, 1 from each subtype. All the training slides were selected in such a way that they are all typical examples of different grading and contain a sufficiently large number of cytologic components of interest. The remaining 33 case images, 10 from undifferentiated, 13 from poorly differentiated, and 10 from differentiating subtypes, were used for testing purposes. With this training and testing data set configuration, the whole slide classification accuracy produced by the computerized system is 87.9%. With this promising high classification accuracy, the system has a good potential to help pathologists in their assessments of grading of differentiation for neuroblastoma with good reproducibility.
No takes yet. Share an insight, caveat, or question.
Kong et al. (2008) studied this question.