Abstracts

RETURN TO ABSTRACT LISTING


2285. Use of an Artificial Intelligence Decision Support Platform for Determining Management of Masses Detected on Screening Breast Ultrasound
Authors * Denotes Presenting Author
  1. Liane Philpotts *; Yale School of Medicine
  2. Joseph Cavallo; Yale School of Medicine
  3. Liva Andrejeva; Yale School of Medicine
  4. Casey Cotton; Yale School of Medicine
  5. Melissa Durand; Yale School of Medicine
  6. Lev Barinov; Koios Medical
Objective:
Incidental solid masses are commonly found on whole breast screening ultrasound (WBUS) the management of which is not well established. The purpose of this study was to retrospectively evaluate the impact of an FDA approved decision support software on BI-RADS 3 and BI-RADS 4 classified lesions detected on screening breast ultrasound.

Materials and Methods:
Following IRB approval, retrospective search of the breast imaging electronic database (PenRad Inc, MN) over a one-year period (10/1/17-9/30/18) was performed to identify lesions reported as BR3 and BR4 on WBUS following a normal digital breast tomosynthesis mammogram. 206 lesions from unique patients (7 malignant and 199 benign) were analyzed via a proprietary artificial intelligence (AI) software platform (Koios DS). The software analyzes orthogonal B-mode images and classifies them according to risk. The system was evaluated against pathological or a minimum of 6-month stable follow-up ground truth. Shifts from BI-RADS 3 and BI-RADS 4 assessments (assigned by dedicated breast radiologists) into BI-RADS 2, 3 and 4 were then evaluated, along with an overall ROC Curve AUC, and sensitivity computed between initial radiologist reads and the system.

Results:
41.1% of benign BI-RADS 3 lesions were downgraded to BI-RADS 2, 26.8% remained as BI-RADS 3 and 32% were upgraded to BI-RADS 4. 2/2 malignant BI-RADS 3 lesions were correctly identified by the system as being suspicious. Evaluating the final distribution of system provided scores, we observed a 45 lesion increase in additional benign biopsy recommendations (p < 1e-10), 102 lesion decrease in benign follow-up p < 1e-5), and a 74 lesion increase in BI-RADS 2 utilization (p < 1e-15). Overall system sensitivity was 100% and radiologist sensitivity was 71.4% (p = 0.127). System ROC curve AUC was 0.89 compared to a Radiologist AUC of 0.79 for an AUC difference of 0.10 (-0.05 ,0.25, 95% CI).

Conclusion:
Utilization of an AI decision support tool for WBUS findings could result in shifts in the final BI-RADS categories with the potential to increase the percent of lesions characterized as benign and increase sensitivity for malignant lesions. Larger and prospective studies in how the tool integrates into the workflow and influences management will be needed.