Classification of colon biopsy images based on novel structural features
Saima Rathore, Muhammad Aksam Iftikhar, Mutawarra Hussain, Abdul Rahman Jalil · 2013
Microscopic analysis of colon biopsy samples is a common medical practice for identifying colon cancer. However, the process is subjective, and leads to significant inter-observer/intra-observer variability. Further, pathologists have to examine many biopsy samples per day, therefore, factors such as expertise and workload of the histopathologists also affect the diagnosis. These limitations of the manual process result in the need of a computer-aided diagnostic system, which can help the histopathologists in accurately determining cancer. Image classification is one of such computer-aided techniques, which can help in efficiently identifying normal and malignant colon biopsy samples without the need of subjective involvement of histopathologists. In this work, we propose a colon biopsy image classification technique, wherein two novel structural feature types, namely, white run-length features and percentage cluster area features have been introduced These features are calculated for each colon biopsy image. The features are reduced using principal component analysis (PCA). The original and the reduced feature sets are then given as input to random forest, rotation forest, and rotation boost classifiers for classification of images into normal and malignant categories. The proposed technique has been evaluated on 174 colon biopsy images, and promising performance has been observed in terms of various well-known performance measures such as accuracy, sensitivity and specificity. The proposed technique has also been proven to be better compared to previously published techniques in the experimental section. Further, performance of the classifiers has been evaluated using ROC curves, and area under the curve (AUC). It has been observed that rotation boost classifier in combination with PCA based feature selection has shown better results in classification compared to other classifiers.