Computer Aided Interpretation of Lesions in Automated 3D Breast Ultrasound

PURPOSE: A CAD system was developed to assist radiologists in characterizing lesions in Automated 3D Breast Ultrasound (ABUS). In this study we investigate if reader performance improves with CAD. METHOD AND MATERIALS: The CAD system computes a lesion malignancy rating when a seed point is placed in the center of the lesion. In this computation features representing spiculation, lesion shape, margin contrast, posterior enhancement, texture, and echogenicity are calculated from a 3D lesion segmentation and combined in a mathematical model using a training database and leave-one-out. Margin contrast and coronal spiculation were the most discriminating features. Images used in this study were obtained from three institutes using the Somo-V system (U-systems, Sunnyvale, CA). 88 patients were included and for each patient one ABUS view with a lesion was selected. 47 lesions were malignant and 41 were benign. Six readers (4 radiologists and 2 residents) read the series of cases two times on a dedicated workstation (HAMAM prototype, Mevis Medical Solutions, Bremen), allowing synchronized viewing of coronal, transversal, and sagittal slices. Readers rated the indicated lesions on a continuous scale ranging from 0 to 100. Readers first read half of the cases with CAD and the other half without. In the second session cases were read with/without CAD if they were first read without/with. Bef ore the study sessions, readers were presented a series of 26 training cases with 23 benign and 14 malignant lesions, to familiarize them with the 3D viewing station and CAD ratings. To compare performance with and without CAD ROC analysis was used (DBM MRMC 2.2). RESULTS: With an area under the curve (AUC) of 0.917 CAD performed very well in comparison to the readers. Without CAD the average AUC of the readers was 0.860 (0.881 for the radiologists, 0.818 for the residents). With CAD the perf ormance of the readers increased to 0.881 (0.886 for the radiologists, 0.869 for the residents). The difference was not significant (p=0.21). CONCLUSION: A CAD system for lesion characterization in ABUS perorms well in comparison to readers. Reader performance did not increase significantly when working with CAD, which may be due to inexperience with ABUS and CAD.