Yanling Chi, Abstract Co-Author: Nothing to Disclose

Sudhakar Kundapur Venkatesh MBBS, FRCR, Abstract Co-Author: Nothing to Disclose

Jiayin Zhou, Abstract Co-Author: Nothing to Disclose

Qi Tian, Abstract Co-Author: Nothing to Disclose

Jimin Liu PhD, Presenter: Nothing to Disclose

A practical computer aided diagnosis is proposed with promising performance.

For a radiologist, detecting small tumors is relatively arduous, since it is time-consuming and small tumors are prone to being overlooked. Visually identifying tumor types is somehow experience dependent. The radiologist with rich experiences can work more efficiently than that with little experiences, but accumulating experiences is a long journey. Measuring tumors, to a certain degree, is subjective. A computer aided diagnosis, which can help the radiologist to improve efficiency in the routine work, is required. We propose to automatically detect once for all liver tumors in the CT scans for the radiologist’s fast identification, and predict the tumor types based on the “past experiences”, a database of tumors with confirmed diagnosis, to support the radiologist’s decision or alert him/her other possibilities.

We proposed to detect liver abnormalities using a subtraction, i.e. normal tissue/organ subtraction, thus, avoid definition of special tumors, and to recognize the focal masses using a similarity query, thus, can be potentially generalize our method to identify any type of tumor if the sample models of this tumor type are available in the database. The goals are fulfilled by a novel context based vascular segmentation method, a novel tumor segmentation method, a novel multiple phase based liver tumor representation method, and a novel hyper-cube based image retrieval method. Our database has 100 clinical CT scans for performance evaluation. Sixty seven scans are at single phase which are from 67 patients and used for tumor detection. The tumor size is of 0.5-180 ml. Thirty three scans have four phases with confirmed diagnosis and are used for tumor classification. They are from 23 patients with four types of tumors: Hepatocellular carcinoma, Metastasis, Hemangioma, and Cyst. The data are with slice thickness of 0.6-5.0 mm, and in-plane resolution of 0.62-0.78mm. We now evaluated eight datasets. The minimal tumor detected is 0.5 ml and the classification accuracy was 90%.

Quantitative measurement, such as the tumor position and size, its distances to the veins etc, can be provided.

Chi, Y, Venkatesh, S, Zhou, J, Tian, Q, Liu, J, A Computer-aided Diagnosis for Liver Focal Mass Detection, Classification, and Quantitative Measurements.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11004682.html