A novel approach for ensemble clustering of colon biopsy images

Colon cancer diagnosis based on microscopic analysis of biopsy sample is a common medical practice. However, the process is subjective, biased and leads to interobserver variability. Further, histopathologists have to analyze many biopsy samples per day. Therefore, factors such as tiredness, experience and workload of histopathologists also affect the diagnosis. These shortcomings require a supporting system, which can help the histopathologists in accurately determining cancer. Image segmentation is one of the techniques, which can help in efficiently segregating colon biopsy image into constituent regions, and accurately localizing the cancer. In this work, we propose a novel colon biopsy image segmentation technique, wherein segmentation has been posed as a classification problem. Local binary patterns (LTP), local ternary patters (LTP), and Haralick features are extracted for each pixel of colon biopsy images. Features are reduced using genetic algorithms and F-Score. Reduced features are given as input to random forest, rotation forest, and rotation boost classifiers for segregation of image into normal, malignant and connecting tissues components. The clustering performance has been evaluated using segmentation accuracy and Davies bouldin index (DBI). Performance of classifiers has also been evaluated using receiver operating characteristics (ROC) curves, and area under the curve (AUC). It is observed that rotation boost in combination with F-Score has shown better results in segmenting the images compared to other classifiers.