Abstract Objective: Ultrasound is one of the most effective non-invasive tool for breast tumor detection and diagnosis. However, acquisition and observation of ultrasound images are mostly subjective and highly dependent on personal experience and judgment. The inter-observer variation often results in significantly different decisions. Objective quantification of sonographic features has become a pressing issue facing the medical staff. This research is to develop a robust, well-performing CAD system which can provide objective suggestion for BI-RADS categorization. Materials & Methods: Sonographic tumor features of 264 tissue-proved BI-RADS 3 to 5 breast lesions (65 cancers, 199 benign cases)were collected consecutively by well-experienced single investigator using Siemens S2000 in National Taiwan University Hospital. There are initially 30 attributes extracted from only one representative view of each tumor. These attributes based on the lexicon descriptions of BI-RADS includes characters such as shape, margin irregularity, heterogeneity, boundary, acoustic shadow, flexibility. After feature selection, 9 attributes, including 6 B-mode attributes and 3 elastography-related attributes are taken into practical model building. All 9 attributes used in this research are quantitative features and are computed by computer. The only necessarily manual effort is to outline the tumor lesion on the ultrasound image by investigator. We randomly take 175 cases(42 cancers, 133 benign cases), about two-third of the sample, for model training and leave 89 cases (23 cancers, 66 benign cases)as the independent test data. In both training data and independent data, the ratio between benignancy and malignancy is kept even. Using “Parameterized Three-Phased Ensemble Model”, we transform the original binary predicting result into probabilistic form so that the model can be applied to BI-RADS category which is based on malignancy probability. In each ensemble model, every component classifier (Multi-Layer FLD Classification Tree) will be assigned a weight according to its performance by AdaBoost algorithm. The classifier gives back a real value as its predicting result. The categorizing probability threshold we adopt in this research follows standard BI-RADS system: BI-RADS 3 <0.02, 4A 0.02∼0.25, 4B 0.26∼0.50, 4C 0.50∼0.89, 5 0.89∼1. Results and Conclusion: The BI-RADS categorizing result on 89 independent cases are as follow: the computer-made proportions of malignancy in each predicted BI-RADS category for 3, 4A, 4B, 4C, 5 are 0.0370, 0.0938, 0.4118, 0.8333 and 1 respectively. On the other hand, the clinician's performance in each BI-RADS category are 0, 0.0909, 0.4166, 0.8333 and 1. The overall performance of CAD auto-categorization is compatible the result made by experienced clinician. This revolutionized approach will give objective evaluation of breast sonographic features. The result will help gives first-line inexperienced physician in patient allocation for referral or follow-up of breast lesions. On the mean time, this system can also provide corresponding second opinion for experienced doctor before making the final decision. Citation Information: Cancer Res 2012;72(24 Suppl):Abstract nr P4-03-02.