Comparisons between artificial intelligence computer-aided detection synthesized mammograms and digital mammograms when used alone and in combination with tomosynthesis images in a virtual screening setting

This retrospective, fully crossed, completely randomized, multireader, multicase study was approved by the Institutional Review Board of our hospital. Written informed consent to was obtained from all the recruited patients enrolled in this study. FUJIFILM (Tokyo, Japan), which invented the new SM for the reading, provided the equipment and support for this study. The data and materials submitted for publication were always under the control of the researchers who were not FUJIFILM personnel. According to the joint research and development agreement, 500 DM + DBT examinations with a bilateral two view (cranio-caudal/mediolateral oblique CC/MLO) image were performed at ? Hospital between January 2020 and August 2020. All DM + DBT examinations were performed with a narrow-angle DBT system with 15-degree tube motion from 15 projection images (AMULET Innovality; FUJIFILM)). DM and DBT were performed during a single breast compression per view. We decided that the comparison of the reader performance was performed at the per-breast level using only mediolateral oblique (MLO) view images taking the reading time and readers’ attention span during the reading study into consideration.

We excluded patients with past surgical history from a case set and developed the case set of 388 breast images from 198 DM/AI CAD SM + DBT examinations in sequence based on the diagnostic work-up patients, which comprises 84 breast cancers, 83 biopsy-proven benign lesions, and 221 normal or benign cases (BI-RADS score of 1 or 2) [10] with negative results after 1 year follow-up. We included dataset with anonymized images and patient’s data with pathology results if an image-guided biopsy and/or surgery was performed. The distribution and mammographic findings were 83 masses (49 malignant and 34 benign), 16 masses with associated calcifications (15 malignant and 1 benign), 5 focal asymmetry densities (2 malignant and 3 benign), 10 architectural distortions (4 malignant and 6 benign), and 53 calcifications (14 malignant and 39 benign). The lesion size of the malignant cases with mass type, ranged from 6 to 42 mm (median, 19 mm; mean, 20 mm). The verified cancer cases included 71 cases of invasive ductal carcinoma, 8 cases of DCIS, 3 cases of invasive lobular carcinoma, 1 case of tubular carcinoma, and 1 case of invasive micropapillary carcinoma. The breast density distribution ratings of the cases were 206 of 388 (53%) non-dense breasts (including almost entirely fatty or scattered fibroglandular densities) and 182 of 388 (47%) dense breasts (including heterogeneously dense or extremely densities), indicating the tissue density BI-RADS score [10]. The age of the study cohort ranged from 23 to 88 years (median, 52 years; mean, 54 years).

We utilized the latest version (under development) of FUJIFILUM image processing approach to generate the AI CAD SM used in this study. The AI algorithm uses a fast and efficient convolutional neural network, ESPNet [11]. They have improved different finding patterns, such as soft tissue densities, linear structures, and calcifications, detected and extracted from DBT data using AI CAD-like technology. They are new pixel-based synthesis algorithm that merges the selected DBT plane based on the AI CAD detections from the DBT data. The lesions detected by AI CAD are not marked or outlined on AI CAD SM.

Four Japanese radiologists who were ranked as having an “A” certification from the Japan Central Organization on Quality Assurance of Breast Cancer Screening (i.e., achieved the scores required for a DM reading instructor in Japan) performed the image evaluations. The radiologists were blinded to the case selection. The readers had between 5 and 30 years of experience in mammography (median, 12 years) and their experience with DBT ranged from 2 to 8 years (median, 4 years). None of the radiologists had prior experience with the new SM reading.

The evaluations of the images by the readers were conducted in two stages. The examinations were read twice as follows: (a) DM followed by DM + DBT, and (b) AI CAD SM followed by AI CAD SM + DBT. At least 4 week intervals were allowed as wash-out periods. In each reading mode, the radiologists were prompted to rate each breast using the two-dimensional (2D) image alone followed by the rating of the DBT using the five-point forced BI-RADS (1, 2, 3, 4, or 5) scores [10]. The most suspicious finding was marked and rated on the electronic reporting system.

The reading times were automatically recorded per examination using the workstation software from the moment the examination was displayed until the reader document the image findings including the locations and radiological characterizations of the cancers and benign lesions were documented. The readers did not use a timer and were blinded to the reading time.

To acquaint themselves with the workstation and interpretation procedure, all radiologists were trained to review sets of 100 DM + DBT and sets of 200 breast images from 100 DM/AI CAD SM + DBT examinations (not included in the study) before participating in this study.

The radiologists used a reading dual-monitor workstation with an electronic reporting system (Climb-Mammography WS; Climb Medical Systems) and a 5MP DB/DBT-certified diagnostic color display (CCL-S500; JVCKENWOOD) calibrated to the DICOM Grayscale Standard Display Function, for the image interpretation.

An unblinded review of every case per breast was performed by one radiologist (?, a breast imaging radiologist with 30 years of experience in mammogram interpretation and 10 years of experience in tomosynthesis image interpretation), who was not involved in the observer study, to determine　the standard pathologic reference and reference mammographic findings. The radiologist also confirmed that there were no discrepancies between the pathologic diagnoses and mammographic findings of 84 breast cancers and 83 biopsy-proven benign lesions. The number of true and false interpretations by the readers was assessed for each image, including the location and radiological characterization of the cancers and benign lesions and the confirmed normal status. This study was modeled in a screening setting: hence, the final assessments of the images were categorized as follows: positive (BI-RADS score, 3–5) and negative (BI-RADS score, 1–2).