Elsevier

Pattern Recognition

Volume 42, Issue 6, June 2009, Pages 1080-1092
Pattern Recognition

Computer-aided evaluation of neuroblastoma on whole-slide histology images: Classifying grade of neuroblastic differentiation

https://doi.org/10.1016/j.patcog.2008.10.035Get rights and content

Abstract

Neuroblastoma (NB) is one of the most frequently occurring cancerous tumors in children. The current grading evaluations for patients with this disease require pathologists to identify certain morphological characteristics with microscopic examinations of tumor tissues. Thanks to the advent of modern digital scanners, it is now feasible to scan cross-section tissue specimens and acquire whole-slide digital images. As a result, computerized analysis of these images can generate key quantifiable parameters and assist pathologists with grading evaluations. In this study, image analysis techniques are applied to histological images of haematoxylin and eosin (H&E) stained slides for identifying image regions associated with different pathological components. Texture features derived from segmented components of tissues are extracted and processed by an automated classifier group trained with sample images with different grades of neuroblastic differentiation in a multi-resolution framework. The trained classification system is tested on 33 whole-slide tumor images. The resulting whole-slide classification accuracy produced by the computerized system is 87.88%. Therefore, the developed system is a promising tool to facilitate grading whole-slide images of NB biopsies with high throughput.

Introduction

Peripheral neuroblastic tumors (pNTs) are a group of embryonal tumors of the sympathetic nervous system, and include neuroblastoma (NB), ganglioneuroblastoma, and ganglioneuroma category [1]. Each year, more than 600 children and adolescents are diagnosed with pNTs in the United States, and it comprises about 8–10% of all childhood cancers [1], [2]. Of all the cancer categories in pNTs, NB, composed of neoplastic neuroblasts in various maturation grades with no or limited Schwannian stromal development, is the most common tumor that affects children ranging from newly born infants to teenagers.

According to the International Neuroblastoma Pathology Classification System (the Shimada system), NB can be further classified into three categories, namely undifferentiated (UD), poorly differentiated (PD), and differentiating (D) subtype, based on the grade of differentiation [3]. A simplified classification tree diagram of this recommended classification system is shown in Fig. 1. NB, with different grades, usually has unique pathological characteristics and micro-texture features [4]. Representative tumors of the three differentiation grades are shown in Fig. 2. Typical features of these subtypes can be briefly summarized as follows:

  • (1)

    Tumors in the UD subtype often present such features as small to medium-sized NB cells, thin cytoplasm, none-to-few neurites, round to elongated nuclei, and the salt and pepper appearance of chromatin with or without prominent nucleoli.

  • (2)

    As for PD cases, the typical rosette formations and/or clearly recognizable neurites are observed in tumor tissues.

  • (3)

    Tumors in the D subtype contain >5% of D neuroblasts characterized by nuclear and cytoplasmic enlargement; an eccentrically located nucleus containing a single prominent nucleolus in most cases; and the increased ratio of the diameter of the cell to that of the nucleus (typically >2).

It is usually the case that the more differentiated tumors are, the less aggressively they behave. As a result, patients with tumors of higher grades of differentiation may have better chances to survive. In clinical practice, treatments for cases with different neuroblastic grades are quite different. For this reason, an accurate grading of a NB sample is crucial to make an appropriate choice of treatment plans.

In current clinical practice, differentiation grading is made with visual examinations of tumors by pathologists under the microscope. There are several weaknesses associated with visual evaluations. First of all, it is often time-consuming and cumbersome for pathologists to review a large number of slides in practice. Secondly, visual evaluations can be subject to unacceptable inter- and even intra-reviewer variations. A recent study reports that there is a 20% discrepancy between central and institutional reviewers [5]. Thirdly, for practical reasons, pathologists often sample slide regions to be examined, making the whole process subject to sampling bias. However, this may lead to erroneous results for tumors exhibiting strong heterogeneity.

To overcome these weaknesses rooted in the visual evaluation process, several computerized methods that automate the image analysis procedures are being developed with promising initial results [6], [7], [8]. However, to the best of our knowledge, no research work, so far, has been devoted to developing a computer-aided classification methodology that automates the process of classifying NB whole-slide images in accordance with the grade of differentiation. In this study, we propose an image analysis framework that integrates intensive computer vision and machine learning techniques for the purpose of grading NB images. Within this system, an image hierarchy consisting of multiple image resolution levels is established for each given tumor image. Furthermore, the system dynamically changes the image resolution level at which it proceeds with sequential image analysis steps. At each image resolution level, every image is first segmented into four cytological components using an automated image segmentation method. Discriminating features extracted from segmented image regions are then used to classify each image into one of the three grading classes by a family of classifiers. The resulting decisions are next combined using a two-step classifier combining mechanism. Each classification decision is first evaluated with a confidence measure that indicates the degree of agreements across different classifiers. Based on the evaluation results, the proposed system either stops its analysis process or continues with further investigations by including more image details.

Section snippets

Image acquisition

In this study, all NB tumor slides are collected from Nationwide Children's Hospital in accordance with an Institutional Review Board (IRB) protocol. According to the protocols commonly used in the Children's Oncology Group, these tissue slides are cut at a thickness of 5μm and soaked in paraffin at the preparation stage. Each NB slide in the dataset is prepared using a dual staining procedure in which haematoxylin and eosin (H&E) are used to increase the visual contrasts among different

Results

Before the developed system can achieve reasonably good grading accuracy, it needs to be well trained. A brief outline of the training–testing process is shown in Fig. 6, where solid and dashed arrows indicate the steps executed online and offline, respectively. In the offline training stage, multiple statistical learning models are used to enrich the system knowledge base with the ground-truth given by an experienced pathologist. As the computer system is trained with more and more typical

Discussions

The developed system is a useful tool that pathologists and clinicians can use in diagnosing the grade of NB. Although decision accuracy of the computerized system is promising in our tests, it should be emphasized that the role of the computerized system should always be limited to that of a second reader or pre-screener who can make consistent decisions on NB classification in parallel with expert pathologists. In other words, the computerized methodology is not intended to replace the

Conclusions

This article presents an automated grading system for the quantitative analysis of the histological images of the H&E stained NB cross-sections. To emulate the way pathologists assess resected specimens, the complete image analysis pipeline is designed within a multi-resolution paradigm. When tested on 33 whole-slide images, the overall classification accuracy of the system is 87.88%. The developed algorithm chooses to work on images of the lowest resolution level where sufficient image details

About the Author—JUN KONG received the B.S. in Information and Control and M.S. in Electrical Engineering from Shanghai Jiao Tong University, Shanghai, China, in 2001 and 2004. He is currently a Ph.D. student in the Department of Electrical and Computer Engineering at Ohio State University, Columbus, OH. His research interests include computer vision, machine learning, and pathological image analysis.

References (25)

  • M.J. Goodman, J.G. Gurney, M.A. Smith, A.E. Olshan, Cancer incidence and survival among children and adolescents:...
  • E.P. Castleberry

    Neuroblastoma

    European Journal of Cancer

    (1997)
  • H. Shimada et al.

    The International Neuroblastoma Pathology Classification (the Shimada system)

    Cancer

    (1999)
  • H. Shimada et al.

    Terminology and morphologic criteria of neuroblastic tumors: recommendation by the International Neuroblastoma Pathology Committee

    Cancer

    (1999)
  • L.A. Teot et al.

    The problems and promise of central pathology review: development of a standardized procedure for the children's oncology group

    Pediatric and Developmental Pathology

    (2007)
  • F.J. Ayres et al.

    Estimation of the tissue composition of the tumor mass in neuroblastoma using segmented CT images

    Medical and Biological Engineering and Computing

    (2004)
  • A. Madabhushi et al.

    Automated detection of prostatic adenocarcinoma from high resolution ex vivo MRI

    IEEE Transactions on Medical Imaging

    (2005)
  • S. Doyle, A. Madabhushi, M.D. Feldman, J.E. Tomaszeweski, A boosting cascade for automated detection of prostate...
  • B.B. Cambazoglu et al.

    Efficient processing of pathological images using the grid: computer-aided prognosis of neuroblastoma

  • Y. Liang et al.

    Multiresolution learning paradigm and signal prediction

    IEEE Transactions on Signal Processing

    (1997)
  • H. Yu, J. Yang, J. Han, Classifying large data sets using SVMs with hierarchical clusters, in: SIGKDD, 2003, pp....
  • Y. He et al.

    A new multiresolution classification model based on partitioning of feature space

    IEEE International Conference on Granular Computing

    (2005)
  • Cited by (141)

    View all citing articles on Scopus

    About the Author—JUN KONG received the B.S. in Information and Control and M.S. in Electrical Engineering from Shanghai Jiao Tong University, Shanghai, China, in 2001 and 2004. He is currently a Ph.D. student in the Department of Electrical and Computer Engineering at Ohio State University, Columbus, OH. His research interests include computer vision, machine learning, and pathological image analysis.

    About the Author—OLCAY SERTEL received his B.S. from Yildiz Technical University, Istanbul, Turkey, and his M.Sc. degree from Yeditepe University, Istanbul, Turkey, in 2004 and 2006, both in Computer Engineering. He is currently a Ph.D. student at the Department of Electrical and Computer Engineering and working as a Graduate Research Associate at the Department of Biomedical Informatics at The Ohio State University, Columbus, OH. His research interests include computer vision, image processing and pattern recognition with applications in medicine.

    About the Author—HIROYUKI SHIMADA is currently an Associate Professor of Clinical in Children's Hospital Los Angeles and Department of Pathology and Laboratory Medicine at The University of Southern California. He is the Pathologist-of-Record for the Children's Cancer Group (CCG) Neuroblastoma Studies. He is also a core member of International Neuroblastoma Pathology Committee and chairing workshops and other activities. In 1999 this Committee established the International Neuroblastoma Pathology Classification (the Shimada system) by adopting my original classification published in 1984. His research interests include investigating morphological characteristics of pediatric tumors.

    About the Author—KIM L. BOYER received the BSEE (with distinction), MSEE, and Ph.D. degrees, all in Electrical Engineering, from Purdue University in 1976, 1977, and 1986, respectively. Since 1986, he has been with the Department of Electrical and Computer Engineering, The Ohio State University, where he holds the rank of a Professor. In 1986, he founded the Signal Analysis and Machine Perception Laboratory at Ohio State. He is a fellow of the IEEE and a fellow of IAPR. He is a member of the Governing Board for the International Association for Pattern Recognition and Chair of the IEEE Computer Society Technical Committee on Pattern Analysis and Machine Intelligence. Dr. Boyer's research interests include all aspects of computer vision and medical image analysis, including perceptual organization, structural analysis, graph theoretical methods, stereopsis in weakly constrained environments, optimal feature extraction, large model bases, and robust methods.

    About the Author—JOEL H. SALTZ is a Professor and Chair of the Department of Biomedical Informatics, Professor in the Department of Computer Science and Engineering at The Ohio State University (OSU), Davis Endowed Chair of Cancer at OSU, and a Senior Fellow of the Ohio Supercomputer Center. He received his M.D. and Ph.D. degrees in Computer Science at Duke University. Dr. Joel Saltz has developed a rich set of middleware optimization and runtime compilation methods that target irregular, adaptive, and multi-resolution applications. Dr. Saltz is also heavily involved in the development of ambitious biomedical applications for high-end computers, very large-scale storage systems, and grid environments. He has played a pioneering role in the development of pathology virtual slide technology and has made major contributions to informatics applications that support point-of-care testing.

    About the Author—METIN N. GURCAN received his M.Sc. degree in Digital Systems Engineering from UMIST, Manchester, England and his Ph.D. degree in Electrical and Electronics Engineering from Bilkent University, Ankara, Turkey. He is working as an Assistant Professor at the Department of Biomedical Informatics at The Ohio State University since 2006. Dr. Gurcan's research interests include image analysis and understanding, computer vision with applications to medicine.

    View full text