Automated morphologic classification of benign and malignant breast masses on ultrasound

Purpose: Benign breast masses can be differentiated from malignant masses based upon morphologic features. However, there has been little exploration of the quantification of these morphologic differences. This study aimed to quantify and classify morphologic characteristics of benign and malignant breast masses on ultrasound imaging.

Methods/Materials: We utilized an existing previously published and publicly available dataset of breast ultrasound images. Study team members quantified the circularity, solidity, and extent of breast masses on ultrasound images. Circularity was defined as the resemblance of the breast mass shape to a circle. Solidity and extent was defined as the extent of the ratios of the breast mass shape to an image processing bound (solidity: convex hull; extent: minimum bounding box). We performed t-test analysis to compare the circularity, solidity, and extent of benign versus malignant masses. Decision tree machine learning analysis of the calculated results was also performed.

Results: A total of 665 breast masses were analyzed (n = 454 benign masses, n = 211 malignant masses). Benign masses had an average circularity of 0.745 while malignant masses showed a circularity of 0.567 (p < .05). Benign masses had an average solidity of 0.980, which was greater than malignant masses which had a solidity of 0.879 (p < .05). Benign masses had an average extent of 0.768, which exceeded the average extent of malignant masses of 0.638 (p < .05). Decision tree machine learning analysis revealed that a circularity of at least 0.762 was 96.64% sensitive for malignancy. A circularity of at least 0.827 was 100% sensitive for malignancy. An extent of at least 0.693 had 75.5% sensitivity and 93.95% specificity for malignancy. A solidity of at least 0.971 was 100% sensitive and 90.2% specific for malignancy.

Conclusions: This study quantified morphologic characteristics of benign and malignant breast masses on ultrasound imaging. We also demonstrated the utility of the automated calculation of these metrics for the classification of benign and malignant masses, particularly solidity, which was shown to be a highly sensitive and specific tool for the detection of malignancy.