Texture Characterization for Whole-slide Histopathological Image Analysis: Applications to Neuroblastoma and Follicular Lymphoma
Olcay Sertel ; The Ohio State University; Jun Kong MSc; The Ohio State University; Gerard Lozanski MD; The Ohio State University; Hiroyuki Shimada MD, PhD; Children's Hospital ; Umit Catalyurek PhD; The Ohio State University; Joel Saltz MD, PhD; The Ohio State University; Metin Gurcan PhD; The Ohio State University;
Content:
Typically the evaluation of histopathology slides is performed by visually inspecting them for certain patterns. This qualitative analysis is time consuming, irreproducible and may lead to inter- and intra-reader variability. Therefore, it is important to develop computerized image analysis systems for quantitative analysis of histopathology slides for reproducible and more consistent cancer prognosis. In this study, we investigate the use of texture characterization for histopathological image analysis with applications to computer-aided prognosis (CAP) systems for neuroblastoma (NB) and follicular lymphoma (FL). NB is a cancer of nerve cell origin and the current classification system is based on several morphological features, such as Schwanian stromal development and grade of neuroblastic differentiation. As part of a complete NB image analysis framework, we propose a computerized system to determine the level of Schwannian stromal development in digitized images of NB slides. FL is a tumor of the lymphatic system and it originates from follicle center B cells comprised of centrocytes and centroblasts. Histological grading of FL is based on the number of centroblasts inside the follicular centers. As the first stage of the CAP system, we propose a computerized system to detect the follicles from the digitized tissue slides. Both stromal development and follicle detection depend on analysis of the underlying image for certain texture patterns.
Technology:
Our system processes images of stained (H&E, CD10, CD20) tissue samples that are digitized at 40x magnification using a whole-slide scanner (Aperio Scanscope, Vista, CA). Due to the limited hardware storage capability, images were compressed following the JPEG compression standards at approximately 1:40 compression ratio. The whole-slide tissue images are typically very large with a spatial resolution up to 80k¿120k with a resulting image size up to 6 GB after being compressed. Therefore we divide the whole-slide image into non-overlapping image tiles and process each image tile separately. Using software to distribute image tiles to different computation nodes of a compute cluster, image analysis routines are applied to each image tile and results are combined for the whole-slide analysis.
Design:
We investigated the use of texture features for different stains in a multi-resolution framework. We also investigated novel ways of combining classical statistical texture features (e.g. contrast, correlation, homogeneity) with novel texture features such as local binary patterns using the Bhattacharya distance. For the NB analysis, we proposed a multi-resolution approach, which mimics the way pathologists evaluate slides. If the classification result obtained at a particular resolution level is not satisfactory, the system automatically switches to higher resolution where the information is more descriptive and hence the classification can be more accurate. The image analysis algorithm for a particular resolution consists of feature extraction, feature selection and classification. For the FL analysis, we extracted textual features to cluster each pixel into two classes: non-follicle or follicle regions and applied a morphological post-processing step to smooth the boundaries of the follicles.
Results:
We provided results that are obtained after processing whole-slide images. We tested our NB classification approach on 43 whole-slide images and 38 of those were correctly classified, leading to a classification accuracy of 88.37%. Moreover, the multi-resolution classification approach provided computational savings up to 60% when compared to regular single scale approach. FL system was evaluated by comparing the automated results with the ground-truth.
Conclusion:
Texture features provide discriminative information for histopathological image analysis. Promising results suggest that the proposed image analysis systems can be used to improve the accuracy, efficiency and reproducibility for automated evaluations of NB and FL in clinical practice.
