Weakly Supervised Multiple Instance Learning for Liver Cirrhosis Classification using Ultrasound Images

We treat the liver cirrhosis classification as a weakly supervised learning problem and propose a Multiple Instance Learning (MIL) method to build a convolutional neural network (CNN) for cirrhosis classification using ultrasound images. In general, regions of interest (ROIs) set carefully in the liver of the ultrasound image are used for feature extraction and classification. Although region placement affects the classification performance, it is difficult to select high quality ROI samples without biopsy because ultrasound images are very noisy. We evaluate our method using liver ultrasound images from 12 cirrhosis patients and 8 healthy individuals that have been diagnosed by clinicians. The experimental results show that the proposed model which is trained by Multiple Instance Learning improved the performance of liver cirrhosis classification using weakly labeled ROIs, compared to the CNN model trained by a conventional learning method. Multiple Instance Learning is an effective method using ROIs placed in liver region of ultrasound images.