Design: In an ongoing IRB-approved clinical study (NCT05245695), enrolling sites were requested to provide imaging, surgical and pathology reports for patients suspected of having endometriosis and scheduled for surgery. Imaging and surgical findings were independently reviewed. Serum protein biomarkers were run in a CLIA-CAP certified lab and processed through a machine learning-based algorithm. These results are from a mid-term analysis of the data.
Setting: Participating sites consisted of 10 women’s health facilities including an academic center, women’s specialty clinics, and fertility clinics.
Patients or Participants: Data from 215 women were collected and used in this study.
Interventions: N/A
Measurements and Main Results: Imaging reports (N=215) identifying confirmed or suspected endometriomas when compared to post-surgery pathology had a sensitivity of 65% (13/20) and specificity of 90.3% (176/195). When imaging reports indicated “confirmed endometrioma,” the sensitivity and specificity for imaging versus pathology was 35% (7/20) and 95.9% (187/195), respectively. In a subset of 74 patients with complete biomarker data, imaging identified endometriomas with a sensitivity of 75% (3/4) and specificity of 91.4% (64/70). When our AI-based algorithm designed to identify endometriomas was added to imaging reports for this dataset, the combined performance could reach a maximum sensitivity of 100% at an appropriate algorithm threshold.
Conclusion: In a limited study, a machine learning-based classifier could enhance sensitivity of imaging detection of endometrioma. This may have application in addition to imaging resources to provide information on surgical preparation or specialty referral.
Pappas, TC*1, Kronfel, CM2, Roy Choudhury, M2, Twiggs, LB2, Haverbusch, V2, Franco, S2, Jarosz, A2. 1Research and Development, Aspira Women's Health, Austin, TX; 2Aspira Women's Health, Austin, TX