LAUSR

Liver cirrhosis is an advanced, diffuse stage of liver injury which usually entails pathologists to check a large number of microscopic images. Obvious differences between liver cirrhosis microscopic images and normal microscopic images, such as the arrangement of hepatocytes, the degree of hepatic fibrosis and the appearance of pseudo lobule, can be efficiently used in medical images classification systems. In this paper, deep learning and standard machine learning methods were applied for helping pathologists making disease diagnosis easier. Firstly, convolutional neural networks and support vector machine were employed to complete the pre-classification of mice liver cirrhosis microscopic images and normal images. We trained the existed convolutional neural networks by our microscopic image datasets after image preprocessing, and we extracted some texture features from all the microscopic images to train the support vector machines; secondly, convolutional neural networks deployed the 98% optimal accuracy that is obviously outperforms support vector machine of 86% final performance. Data augmentation is an efficient approach for solving the problem of insufficient image number. Finally, in experiments, the classification results after data augmentation are more accurate and the trained models are more stable. Moreover, more samples need to be obtained to train the used convolutional neural networks and more features also need to be extracted that are critical to diagnose for pathologists in future works.