Clinical Value Of Deep Learning-based Mammography Diagnostic System For Chinese Women

PARTⅠ Clinical value of a deep learning-based mammography density classification systemObjective To investigate the clinical value of our deep learning（DL）system for mammographic breast density classification.Materials and Methods We analyzed 1686 and 2046 mammographic cases without and with the assistance from the DL system,respectively.Five breast imaging radiologists were divided into two groups: the junior group,in which radiologists had less than ten years of experience,and the senior group,in which radiologists had more than ten years of experience.They conducted binary classification（dense versus non-dense breast）and four-category classification（almost entirely fatty,scattered areas of fibroglandular tissue,heterogeneously dense,and extremely dense）on all mammograms.The accuracy and clinical acceptance of the DL system’s assessment were measured.Linear-weighted kappa（K）and intra-class correlation coefficient（ICC）statistics were applied to analyze and compare the consistency between not only radiologists in the two groups but also all radiologists with or without the assistance of the DL system.Results The accuracy and clinical acceptance of the DL system for the junior group were96.3% and 96.8% for two-category evaluation,and 85.6% and 89.6% for four-category evaluation,respectively.For the senior group,the accuracy and clinical acceptance were 95.5%and 98.0% for two-category evaluation,and 84.3% and 95.3% for four-category evaluation,respectively.The consistency within the junior group,the senior group,and among all radiologists improved with the help of the DL system.For two-category evaluation,their K and ICC values improved to 0.811,0.812,and 0.795 from 0.729,0.749,and 0.760.And for four categories,their K and ICC values improved to 0.810,0.817,and 0.820 from 0.729,0.794,and0.778,respectively.Conclusion The DL system showed high classification accuracy and clinical acceptance in breast density categories.It is helpful to improve radiologists’ consistency on mammographic breast density classification.PART Ⅱ Clinical value of a deep learning-based mammography lesion detection systemObjective To investigate the clinical value of our deep learning（DL）system for mammography lesion detection.Materials and Methods We retrospectively collected 1755 screening and diagnostic mammography cases,each taken in Craniocaudal view（CC view）and Mediolateral Oblique view（MLO view）.The DL system and two junior radiologists,who had less than ten years of experience,blindly read the mammograms individually and independently.Using the review result from a senior radiologist,who had more than twenty years of experience,as a reference standard,we compare the sensitivity for the detection of calcification and mass from the DL system and two junior radiologists.Chi-square test was used to analyze the influence of different factors like patient age,mammography density categories,BI-RADS categories,calcification morphology,calcification distribution,mass morphology,mass size,mass margin and mass density,on the DL system and two junior radiologists.Results A total of 1755 mammography cases were included in this study,of which 885 cases of calcification were detected（a total of 3067 calcifications）,324 cases of mass were detected（a total of 618 masses）.⑴The sensitivity of the DL system and the two junior radiologists to the detection of the calcification was 97.52%（2991/3067）,94.33%（2893/3067）,92.53%（2838/3067）,respectively.The DL system and two junior radiologists had statistically significant differences in the detection of different calcification distributions（P<0.05）.The DL system showed no statistically significant difference in the calcification detection of different patients’ age,mammographic density categories,calcification morphology,and BI-RADS categories（P>0.05）,while the difference between two junior radiologists was statistically significant（P<0.05）.⑵The sensitivity of the DL system and the two junior radiologists to the detection of mass was 86.25%（533/618）,83.98%（519/618）,77.99%（482/618）,respectively.The DL system showed no statistically significant difference in the mass detection of different mammographic density categories（P>0.05）,while the difference between the two junior radiologists was statistically significant（P<0.05）.However,the DL system and two junior radiologists had statistically significant differences in the detection of different patients’ age,mass morphology,size,margin,density and BI-RADS categories（P<0.05）.Conclusion The DL-based mammography lesion detection system has a high sensitivity for the detection of calcification and mass lesions.It has certain stability for the detection of calcification,which can assist radiologists to reduce missed diagnosis,especially suspicious calcification.And DL system is not affected by mammographic density,which can help radiologists reduce the missed diagnosis of masses caused by dense fibroglandular tissue.PART Ⅲ Clinical value of a deep learning-based mammography lesion detection system for BI-RADS 0 lesionsObjective To explore the clinical value of a deep learning-based mammography lesion detection system for BI-RADS 0 lesions.Materials and Methods We retrospectively collected 1075 cases of BI-RADS category 0patients.Among them,997 cases of unilateral and 78 cases of bilateral breasts were judged as BI-RADS category 0.We recorded each patient’s age,breast density category,the reason for BI-RADS category 0,further imaging examination and follow-up results.Four junior radiologists used a double-blind and crossover method to read 80 randomly selected BI-RADS category 0 cases,without and with the assistance of the DL system.We analyzed the junior radiologists’ diagnosis accuracy of BI-RADS category 0 lesions and the time for reading mammograms.Then,two senior radiologists read the 1075 BI-RADS category 0 cases with the assistance of the DL system.We compared the detection rate of BI-RADS category 0 lesions between the DL system and senior radiologists + DL system.Results Among the 80 cases randomly selected from the data set,the consistency and accuracy of the four junior radiologists were poor.In 40 cases,junior radiologists’ sensitivity was 0.18 and0.36,respectively,and the longest reading time was 82 minutes.For comparison,the sensitivity of the DL system was 0.41,and the reading time was only 10 minutes.The sensitivity of the junior radiologists + DL system increased to 0.45,and the reading time was shortened accordingly.In other 40 cases,the sensitivity of the junior radiologist + DL system increased to0.69.Four junior radiologists had poor consistency,with a K value of 0.2（P>0.05）.In addition,two senior radiologists + DL system has a high diagnostic accuracy rate for BI-RADS category 0lesions,among which the highest sensitivity was 0.89,specificity was 0.58,positive predictive value was 0.61,negative predictive value was 0.88,and the AUC value was 0.73.Conclusion The DL system helps to improve the accuracy of the diagnosis of BI-RADS categories 0 lesions by junior radiologists and shorten the reading time.The combination of the DL system with senior radiologists has a higher diagnostic accuracy for BI-RADS categories 0lesions.It can reduce unnecessary recalls and medical costs.