Breast cancers missed during screening in a tertiary-care hospital mammography facility

Impact of radiomics on the breast ultrasound radiologist's clinical practice: From lumpologist to data wrangler

OBJECTIVE

The study aims to assess the impact of radiomics in the clinical practice of breast ultrasound, to determine which lesions are undetermined by the software, and to discuss the future of the radiologist's role.

METHODS

Consecutive analyses of 207 ultrasound masses from January 2018 to April 2019 referred for percutaneous breast biopsy. Breast masses were classified using dedicated ultrasound software (AI). The AI software automatically classified the masses on a scale of 0-100, where 100 is the most suspicious. We adopt the histology results as the gold standard. The cut-off point of malignancy by radiomics was determined, with ±10 % of margin error according to the Youden's index. We considered these lesions as undetermined masses. The performance of the AI software and the radiologist classification was compared using the area under roc curves (AUROC). We also discuss the impact of radiologist validation of AI results, especially in undetermined lesions.

RESULTS

Of the 207 evaluated masses, 143 were benign, and 64 were malignant. The Youden's index was 0.516, including undetermined masses with a varied range of 10 % (0.464-0.567). Twenty-one (14.58 %) benign and twelve (19.05 %) malignant masses were in this range. The best accuracy performance to classify masses was the combination of the reader and AI (0.829). The most common undetermined masses in AI were fibroadenoma, followed by phyllodes tumor, steatonecrosis as benign. Whereas, low-grade, and high-grade invasive ductal carcinoma represents the malignant lesions.

CONCLUSIONS

Artificial Intelligence has a reliable performance in ultrasound breast masses classification. Radiologist validation is critical to determine the final BI-RADS assessment, especially in undetermined masses to obtain the best classification performance.