Motivation for use of AI system with DBT

• Although use of DBT improves cancer detection and recall rates [1-4]

• Reading times are about twice as long as reading digital mammography alone [1, 5, 6]

AI system description

• Developed for detection of malignant masses, architectural distortions, asymmetries and calcifications from reconstructed DBT volumes

• Detection algorithm is based on deep convolutional neural networks that localize, segment and classify breast lesions utilizing only the image pixel data

Algorithm provides (Figures 1 and 2)

• Contour outlines of detected lesions (findings) in DBT slices

• Certainty of Finding Scores: calibrated lesion-level 0-100% scores for each of the identified locations representing the algorithm’s confidence that the detection is malignant

• Case Score: calibrated case-level 0-100% score that is computed by aggregating information from all detections within the case to represent the algorithm’s confidence that the case is malignant

Fig. 1: 61-yo with screening DBT demonstrating a spiculated mass in right anterior lateral breast shown to be a 1.3-cm invasive ductal carcinoma with DCIS (ER+, PR-, HER2-)

Fig. 2: Zoomed views of AI lesion outlines in DBT slices with lesion-level scores and case-level score

Multi-reader, multi-case (MRMC) reader study designed to assess effectiveness and efficiency of concurrent use of AI system

• Conducted with 24 radiologists each reading 260 cases (65 cancer and 195 non-cancer) both with and without AI

          -   Non-cancer cases included 109 negative cases, 21 recalled cases

              shown to be benign without biopsy warranted based on additional

              imaging and 65 biopsy-proven benign cases

          -   Cases retrospectively collected with HIPAA compliance under IRB

              approval with waiver of informed consent

• Minimum 4-week washout period between readings of same case with and without AI (Figure 3)

• Results showed statistically significant improvements in area under the ROC curve (AUC), sensitivity, specificity, recall rate and reading time (Figure 4)

Fig. 3: Multi-reader, multi-case (MRMC) reader study design

Fig. 4: Effects on radiologist performance and reading time from concurrent use of the AI system