Multi-Instance Image Classification for Cancer Diagnosis using Statistical Mapping Support Vector Machine

This paper presents multi-instance (MI) image classification for cancer diagnosis using statistical mapping Support Vector Machine (SVM). The existing MI image classification is limited to focusing on standard multi-instance classification (MIC) assumption, but do not generalize to the whole range of MI data and do not fully utilize the power of conventional SVM. The standard MIC assumption labelled a bag of image as positive if there is at least one instance in it which is positive. Unfortunately, this assumption is not applicable if there is less information about abnormal instances provided in a bag. Therefore, the paper aims to propose conventional SVM that utilized the basic statistical mapping to form a bag vector of instances in order to classify MI images and give the benefit of the automated image diagnostic procedure. Numerical tests examine the benefit of instances’ features transformation to be a vector of bag representation using mean and covariance mapping to Linear-SVM, Square-SVM and Cube-SVM. The experiments used a secondary dataset. The numerical dataset extracted breast histopathology image of 58 patients, which contains 708 features and 2002 instances. The result obtained shows that the proposed SVM can achieve 100% sensitivity after utilizing the covariance mapping with Square-SVM. It means the classification task able to detect the malignant class. In conclusion, the conventional SVM has great potential to improve medical diagnostic procedure using MI image, particularly for cancer diagnostic after adapting statistical features transformation.