Early diagnosis of breast cancer possesses a profound clinical significance in improving the survival rate while reducing the rate of mortality [9, 16]. Ultrasound is one of the commonly used methods for imaging breast lesions [17]. It is convenient, radiation-free and effectively identifies benign and malignant breast lesions [18]. Yet, it is generally necessary to adjust the scanning pressure, focus area and gain in real time to properly analyze the characteristics of the lesion. Meanwhile, the examination of ultrasonography relies on the physician experience and is prone to lead to a missed diagnosis. Furthermore, diagnosing a mass less than 1 cm3 is laborious. AI breast is the innovative function of the new generation eL18-4 probe, which can mark the scanning range, automatically identifying the lesion location, improving the ability to locate nodules.
In this study, a substantial agreement for the scan time in two methods (ICC = 0.7642) was found. The average time to complete a bilateral whole-breast US examination between the three readers was 6.663, 7.686 and 6.083, which cannot generate a clinical report immediately. Scan time correlates with the number of lesions. Simultaneously, there was a substantial agreement for the scan time description between the two readers using AI-breast scan method (ICC = 0.7971). The scan time agreement for the two readers using AI breast US is higher than conventional US method, meaning that the AI breast US helps save time, improving departmental efficiency and integrating into the workflow in a busy breast imaging center using virtual mapping.
The substantial agreement for the number of nodules (ICC = 0.7692) between HHUS and AI-breast is valid. Meanwhile, substantial agreement for the number of nodules description between the two readers of AI-breast (ICC = 0.7843) was observed. In this study, we found a substantial moderate consistency was observed between HHUS and AI breast for more than 6 nodules and a substantial consistency between the two readers for AI breast. In the group with more than 10 nodules, AI breast US procedure identified more nodules. This is because, when the number of nodules higher than 6 it is difficult, even for experienced radiologist, to label each nodule using conventional ultrasound. Moreover, AI breast US can mark a lesion in real time in the energetic body to avoid this complication. Radiologists may also miss some breast tissue. While the AI breast US procedure is assigned a color to demonstrate it has been scanned in one direction on-screen, then assigned a different color once scanned in the opposite plane. This helps reduce the operational dependency and subjectivity during exams.
In this study, we also summarized an perfect conclusion for the location of nodules, including clock face location, and distance to the nipple between a resident physician and an experienced radiologist. There was a uniformly perfect agreement for the clock face location, and distance to the nipple between the two readers using the AI breast procedure. This difference between the AI breast two readers is consistent with the coherence between HHUS and AI-breast, where the latter can auto-annotate the lesion with the position and distance from the nipple for the resident physician, allowing for the elimination of labeling errors and enhance reproducibility. This is particularly valuable for the current diagnostic and follow-up examinations when multiple lesions are present.
Breast cancer diagnosis using ultrasound is based on the image features. In this study, a substantial agreement was observed for the lesion features such as boundary, morphology, aspect ratio, and internal echo (k of 0.7424, 0.7217, 0.6741, 0.6419) between both methods. Moreover, substantial agreement between the two readers of AI breast was observed for the same lesion features (k of 0.6531, 0.6762, 0.6439, 0.6137) were similarly moderated. The k value for both methods conveyed consistent results probably due to simple identification in view of a standard acquisition of lesion in AI breast US by resident physicians. AI Breast is capable of not only recording the transducer position and orientation for each acquired lesion, but it can dynamically indicate areas of the breast scanned after the examination.
Based on the comprehensive description of the images above, the consistency evaluation (Kappa) of BI-RADS classification group for lesions by the radiologist and resident physicians connecting AI breast US scan with conventional US scan method was performed. We found a moderate agreement on lesion management, with a k of 0.6241 for BI-RADS final assessments between the two methods and the two readers of AI breast. Previous studies have also reported a moderate agreement in breast US with mammography. This may be because the AI breast US cannot automatically assess the final BI-RADS category. At same time, it was confirmed that the method has excellent repeatability and could reduce the operational dependency and subjectivity during examinations to avoid missed diagnosis.
There were several limitations in our study, such as the long-term follow-up data on the prognoses of these patients, which are still required. Ample and multicenter studies stemming from this study may provide further information for a larger number of patients. Furthermore, the images may be affected by uncertainties and inaccuracies for requiring specific fuzzy image pre-processing steps [19,20,21,22].
In summary, AI breast US is less operator-dependent due to the precise documentation of lesion location and distance from the nipple and annotation. In the future, the scanning method can be used to monitor benign lesions since the operator can perform AI breast US examination without identifying the lesion.