Many visceral organs contain two important structures: capsule and parenchyma. As a non-invasive examination method, ultrasound is widely used. In this study, we develop a computational framework that consists of capsule localization and parenchyma assessment for disease diagnosis.Clinical decision support system helps providing standard, objective and timely diagnosis. However, although current cirrhosis diagnosis methods are mainly based on the images produced from medical imaging technique and experienced clinicians subjective analysis, many imaging techniques are invasive, and experienced clinicians are in short supply, especially in underdeveloped regions. The proposed system employs an incremental classification model for grading parenchyma patch stages in predicting cirrhosis stages from high frequency ultrasound images. The incremental classification model is based on auto-extracted membrane structures along with patch-ensemble model, using a severe-first strategy. The framework firstly applies multi-scale capsule extraction automatically. The lesions of the capsule and parenchyma show increasing changes in the first three stages. In the final stage, due to the liver ascites, partial lesions restore to normal. To handle the inconsecutive change, the proposed patch-ensemble model applies two layers. We firstly recognize images belong to the severe stage via capsule-specific CIFAR in the first layer. Parenchyma-aimed Resnet is then applied to classify rest images into rest stages. In each layer a data split and aggregation scheme is proposed to evaluate the cirrhosis stage for liver images.The experimental results demonstrate that the proposed method achieves high precision and effectiveness and can be effectively applied to the auxiliary diagnosis of cirrhosis. Some integral parts of the system are also available for visceral organs with similar structure.