Efficiency of mammography detection system based on deep learning for breast suspicious calcifications / 中国医学影像技术

ABSTRACT

Objective: To observe the efficiency of mammography detection system based on deep learning (DL) for breast suspicious calcifications. Methods: Standard cranio-caudal (CC) and medio-lateral oblique (MLO) view breast images of 932 women were interpreted by two junior radiologists and the DL system, respectively. Then the results were reviewed by a senior radiologist. The sensitivities of the two junior radiologists and DL model were compared. Then two-way tests were used to annalyze differences under different BI-RADS categories, morphologies and distributions. Results: There were 274 suspicious calcifications in the ground truth from 932 cases (3 728 images). The sensitivity of two junior radiologists and DL system was 76.64% (210/274), 82.12% (225/274) and 99.64% (273/274), respectively. No significant difference of the DL system was found under different morphologies, distributions and BI-RADS categories (all P>0.05), while for the junior radiologists, the sensitivities of amorphous or grouped calcifications were significantly lower than of the others (both P<0.05). Conclusion: The automatic mammography suspicious calcification detection system based on DL show promising sensitivities and robustness, which may help radiologists to reduce the missing of suspicious calcifications.