2106: Comparison of 3D and 2D ultrasound for differentiating benign and malignant breast masses

Abstract

Objectives: To evaluate the performance of radiologists and computer-aided diagnosis (CAD) systems on 3D US and conventional 2D US for the characterization of solid breast masses.Methods: One hundred fifty solid breast masses (61 cancers and 89 benign) were collected using both 2D and 3D US (Voluson 530D, GE-Kretz). Six radiologists independently and blindly reviewed static 2D US images and stored 3D volume data with a four-week interval and provided a final assessment category to indicate the probability of malignancy. For a CAD study, the extracted texture features of the 2D and 3D US images were used to classify the tumor as benign or malignant using the artificial neural network.Results: For all radiologists, 3D US was superior to 2D US in sensitivity (97.8 ± 2.8 versus 95.8 ± 3.9), specificity (69.5 ± 15.2 versus 64.5 ± 16.8) and negative predictive (98.5 ± 2.1 versus 96.5 ± 3.0) values (p = 0.03 in sensitivity). At ROC analysis, mean Az values of 3D and 2D US by six radiologists were 0.90 and 0.85, respectively. 3D and 2D CAD schemes yielded Az values of 0.97 and 0.85 in distinguishing between benign and malignant lesions, respectively.Conclusions: Stored 3D US information may allow radiologists and CAD systems to do better classification of solid breast masses than is possible with current 2D US images. Objectives: To evaluate the performance of radiologists and computer-aided diagnosis (CAD) systems on 3D US and conventional 2D US for the characterization of solid breast masses. Methods: One hundred fifty solid breast masses (61 cancers and 89 benign) were collected using both 2D and 3D US (Voluson 530D, GE-Kretz). Six radiologists independently and blindly reviewed static 2D US images and stored 3D volume data with a four-week interval and provided a final assessment category to indicate the probability of malignancy. For a CAD study, the extracted texture features of the 2D and 3D US images were used to classify the tumor as benign or malignant using the artificial neural network. Results: For all radiologists, 3D US was superior to 2D US in sensitivity (97.8 ± 2.8 versus 95.8 ± 3.9), specificity (69.5 ± 15.2 versus 64.5 ± 16.8) and negative predictive (98.5 ± 2.1 versus 96.5 ± 3.0) values (p = 0.03 in sensitivity). At ROC analysis, mean Az values of 3D and 2D US by six radiologists were 0.90 and 0.85, respectively. 3D and 2D CAD schemes yielded Az values of 0.97 and 0.85 in distinguishing between benign and malignant lesions, respectively. Conclusions: Stored 3D US information may allow radiologists and CAD systems to do better classification of solid breast masses than is possible with current 2D US images.