Analysis of Malignancy Using Enhanced GraphCut-Based Clustering for Diagnosis of Bone Cancer

Abstract

Osteosarcoma and Ewing’s sarcoma are very common bone tumors, and its biopsy is characterized with spatial distributions of osteoblasts, osteocytes, and osteoclasts. Any abnormal growth found in these three cells can be either cancerous or benign. This paper presents enhanced GraphCut-based clustering framework to ascertain malignancy level in hematoxylin and eosin (H&E)-stained histopathological images. This approach executes iterative GraphCut method to extract foreground objects from biopsy image. Usually, iterative GraphCut needs user interaction to initialize segmentation process. But in enhanced GraphCut method, this initial data is manually generated using standard image processing tools. By doing this, experiment shows that quality of proposed segmentation result is improved. After segmentation of all tissue cells, its categorization is done through color and topological characteristics. Therefore, domain-specific methods such as color-based clustering, mathematical morphology, and active contour are used for feature extraction. This computed features are used to quantify the characteristics of malignancy and classify them as normal, benign, and cancerous using multiclass random forest framework. Proposed method is compared with earlier methods which yields 90% of classification accuracy.