Clinical Image Analysis and Neuroblastoma

• Segmentation of histopathological images
• Neuroblastoma
• Grade of Differentiation
• Stroma Classification

Segmentation of histopathological images

Extensive research is being done with regard to the segmentation of histopathological images. Prognosis classification of neuroblastoma is based on observing morphological cell characteristics. A method had been developed for H&E-stained neuroblastoma images based on morphological top-hat reconstruction. The algorithm developed is used in conjunction with hysteresis thresholding, which was demonstrated to be more accurate than constant thresholding. The nuclei are detected, delineated, and ultimately segmented. This method has an average accuracy of 90.24±5.14%, when compared to manual segmentation.

Neuroblastoma

A computerized pathological image analysis system has been developed to classify Schwannian stromal development and grade of differentiation for neuroblastoma. The system was placed on a cluster of computers with automated load balancing, which significantly reduced run time. The use of a multi-resolution further reduced run time. Stromal classification accuracy was 96.6%, while the grade of differentiation accuracy was 95.6%.

Grade of Differentiation

An automated process to determine the grade of differentiation in neuroblastoma had been developed. The process is employed in a multi-resolution framework. Segmentation is done using the Fisher-Rao criteria embedded in the generic Expectation-Maximization algorithm. Classification is done in two steps, consisting of a majority voting process and a weighted sum rule using priori classifier accuracies. The system had the best overall accuracy of 96.89% when tested. The use of multi-resolution in conjunction with the automated feature selection process led to 34% computational savings. This promising method may allow for efficient classification of neuroblastoma in the future.

In addition, research has been done on the effect of application-level parameters on the performance and accuracy of the system. A number of execution strategies have been developed to allow the system to dynamically choose parameter values, which allows it to trade off performance for accuracy or vice-versa to meet user requirements. Images are processed at varying resolutions, with low resolutions improving performance at the cost of accuracy. This allows for an analysis that reflects the necessities of the user.

Stroma Classification

Currently, analysis of neuroblastoma by pathologists can be a slow, tedious process, with significant reader variability. Thus, an automated computerized system has been developed to analyze neuroblastoma images and classify them as stroma-rich or stroma-poor. The test set was classified with an accuracy of 88.37%. The use of a multi-resolution approach resulted in 85% computational savings. Further refinement of the system may allow it to be used in clinical practice.

Project Researchers

Metin Gurcan, Ph. D.
Joel Saltz, M.D., Ph.D.
Olcay Sertel

Project Publications