Sun Yang Chung MD, Abstract Co-Author: Nothing to Disclose

Woo Kyung Moon MD, Presenter: Nothing to Disclose

Nariya Cho MD, Abstract Co-Author: Nothing to Disclose

Sin Ho Lee MD, Abstract Co-Author: Nothing to Disclose

Yu-Len Huang, Abstract Co-Author: Nothing to Disclose

Dar-Ren Chen, Abstract Co-Author: Nothing to Disclose

The computer-aided diagnosis (CAD) system for classification of breast tumors on ultrasound (US) is under development. However, most of the strategies were performed in a specific US machine and the results were not applicable for different US systems. This study was aimed to develop a CAD system with textural features and image retrieval techniques to classify benign and malignant breast tumors on different US systems.

We evaluated a series of histologically proven 600 solid breast nodules (230 malignant and 370 benign tumors) with four different US systems (Aloka SDD 1200, GE LOGIQ 700, ATL HDI 3000 and HDI 5000 scanners). The suspicious tumor region in the US image was manually selected as region-of-interest subimages and co-variance texture parameters were used in the diagnosis system. This study employed the principal component analysis to project the original textural features into a lower dimensional principal vector that captured most of the textural information. The image retrieval techniques were utilized to differentiate breast tumors based on the similarity of the principal vectors. The k-fold cross-validation method and receiver operator characteristic curve analysis was to estimate the performance of the proposed CAD system.

The accuracy, sensitivity, specificity, and positive and negative predictive values of US classification based on the image retrieval technique was 91.1% (547 of 600), 97.0% (223 of 230), 87.6% (324 of 370), 82.9% (223 of 269), and 97.9% (324 of 331) at RN9 (RNk denotes that the number of retrieving k different US images from the US image database for diagnosing breast tumors). In the four US systems, the receiver operating characteristic area index (Az) for the proposed CAD system was 0.97, 0.91, 0.94 and 0.98 with the mean Az of 0.97±0.01.

The CAD system classified breast tumors seen at four different US systems as benign and malignant with a high accuracy. Because historical cases are directly added into the database, the retraining procedure can be avoided in the proposed system.

Chung, S, Moon, W, Cho, N, Lee, S, Huang, Y, Chen, D, Image Retrieval and Principal Component Analysis for Computer-aided Diagnosis (CAD) of Breast Tumors on Different Ultrasound Systems.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4415214.html