Statistical texture feature set based classification of histopathological images of stomach adenocarcinoma

Inspection of the biopsy samples microscopically plays a vital role in the definitive diagnosis of cancer. To overcome the subjectivity in pathologists’ decision, objective analysis of the stomach biopsy samples is carried out in this work. At the tissue level, malignancy leads to distortions in glandular structure and nearby supporting tissue namely, stroma. These pathological alterations are known to cause larger variations in the image’s texture. So the proposed method extracts textural features from the histopathological image and classifies them using the SVM classifier. Gray-Tone spatial Dependence Matrix (GSDM), Gray-Level Run Length Matrix (GLRLM), Wavelet transform were used for the extraction of statistical texture features from Region of Interest (ROI). Embedded model of feature selection was carried out. Minimum Redundancy Maximum Relevance (MRMR) scheme was used for Feature ranking. The composite feature set comprising of both the spatial domain (GSDM, GLRLM) and Wavelet domain features showed better discriminating characteristics for the four different classes, (Normal, Well Differentiated, Moderately Differentiated, Poorly Differentiated) achieving a highest classification accuracy of 93.75%.