We conducted a multicentric real-world study by including patients who performed digital mammography (DM) or synthetic 2D-views from digital breast tomosynthesis (s2D-views) at 3 centers of Centro Diagnostico Italiano (Milan, Italy). DICOM images of these patients were sent automatically through secure DICOM-communication protocol into the central PACS (Picture Archiving and Communication System) of the center.

The Trace4BDensity™ deep learning tool (DeepTrace Technologies, Italy) was installed on a server of the center as an on-premise service and integrated with a PACS research node. The Trace4BDensity™ tool had previously showed a substantial reliability with board-certified radiologists in dense (A+B) versus nondense (C+D) categories (see Magni et al., 2022), with a Cohen k of 0.807.

The service was implemented with the aim of automatically performing DICOM queries to eligible DMs/s2D-views for assigning BI-RADS breast density classification. The resulting classification was then automatically returned as a DICOM-encapsulated report within the PACS research node, also explaining the probability of the breast density category proposed by the model.

Statistics on usage, failures, classification distribution into the 4 breast density categories, and reliability with expert human readers were calculated for 10 days of continuous functioning.